Adding upstream version 124.0.1.upstream/124.0.1

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

36 files changed, 66491 insertions, 0 deletions

diff --git a/media/libaom/config/aom_version.h b/media/libaom/config/aom_version.h
new file mode 100644
index 0000000000..b791ef64dd
--- /dev/null
+++ b/media/libaom/config/aom_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  0
+#define VERSION_MINOR  1
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v0.1.0"
+#define VERSION_STRING      " v0.1.0"
diff --git a/media/libaom/config/generic/config/aom_config.asm b/media/libaom/config/generic/config/aom_config.asm
new file mode 100644
index 0000000000..be0715562c
--- /dev/null
+++ b/media/libaom/config/generic/config/aom_config.asm
@@ -0,0 +1,92 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+AOM_ARCH_AARCH64 equ 0
+AOM_ARCH_ARM equ 0
+AOM_ARCH_PPC equ 0
+AOM_ARCH_X86 equ 0
+AOM_ARCH_X86_64 equ 0
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_HIGHBITDEPTH equ 1
+CONFIG_AV1_TEMPORAL_DENOISING equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITRATE_ACCURACY equ 0
+CONFIG_BITRATE_ACCURACY_BL equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_COMPONENT_TIMING equ 0
+CONFIG_COLLECT_PARTITION_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_CWG_C013 equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_EXCLUDE_SIMD_MISMATCH equ 0
+CONFIG_FPMT_TEST equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBVMAF_PSNR_PEAK equ 1
+CONFIG_LIBYUV equ 0
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NN_V2 equ 0
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OPTICAL_FLOW_API equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_OUTPUT_FRAME_SIZE equ 0
+CONFIG_PARTITION_SEARCH_ORDER equ 0
+CONFIG_PIC equ 0
+CONFIG_RATECTRL_LOG equ 0
+CONFIG_RD_COMMAND equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_RT_ML_PARTITIONING equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SALIENCY_MAP equ 0
+CONFIG_SHARED equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_SPEED_STATS equ 0
+CONFIG_TFLITE equ 0
+CONFIG_THREE_PASS equ 0
+CONFIG_TUNE_BUTTERAUGLI equ 0
+CONFIG_TUNE_VMAF equ 0
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+FORCE_HIGHBITDEPTH_DECODING equ 0
+HAVE_ARM_CRC32 equ 0
+HAVE_AVX equ 0
+HAVE_AVX2 equ 0
+HAVE_FEXCEPT equ 1
+HAVE_MMX equ 0
+HAVE_NEON equ 0
+HAVE_NEON_DOTPROD equ 0
+HAVE_NEON_I8MM equ 0
+HAVE_SSE equ 0
+HAVE_SSE2 equ 0
+HAVE_SSE3 equ 0
+HAVE_SSE4_1 equ 0
+HAVE_SSE4_2 equ 0
+HAVE_SSSE3 equ 0
+HAVE_SVE equ 0
+HAVE_VSX equ 0
+HAVE_WXWIDGETS equ 0
+STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/generic/config/aom_config.h b/media/libaom/config/generic/config/aom_config.h
new file mode 100644
index 0000000000..a695b0b3e6
--- /dev/null
+++ b/media/libaom/config/generic/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 0
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 0
+#define AOM_ARCH_X86_64 0
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 0
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 0
+#define HAVE_NEON 0
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_SSE4_2 0
+#define HAVE_SSSE3 0
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/generic/config/aom_dsp_rtcd.h b/media/libaom/config/generic/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..0418b3568e
--- /dev/null
+++ b/media/libaom/config/generic/config/aom_dsp_rtcd.h
@@ -0,0 +1,4713 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+#define aom_avg_4x4 aom_avg_4x4_c
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+#define aom_avg_8x8 aom_avg_8x8_c
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+#define aom_avg_8x8_quad aom_avg_8x8_quad_c
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+#define aom_blend_a64_hmask aom_blend_a64_hmask_c
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+#define aom_blend_a64_mask aom_blend_a64_mask_c
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+#define aom_blend_a64_vmask aom_blend_a64_vmask_c
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_pred aom_comp_avg_pred_c
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+#define aom_comp_mask_pred aom_comp_mask_pred_c
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+#define aom_compute_flow_at_point aom_compute_flow_at_point_c
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve8_horiz aom_convolve8_horiz_c
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve8_vert aom_convolve8_vert_c
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+#define aom_convolve_copy aom_convolve_copy_c
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_c
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_c
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_c
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x32 aom_dc_128_predictor_32x32_c
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x64 aom_dc_128_predictor_32x64_c
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_c
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_c
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_c
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_64x16 aom_dc_128_predictor_64x16_c
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_64x32 aom_dc_128_predictor_64x32_c
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_64x64 aom_dc_128_predictor_64x64_c
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_c
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_c
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_c
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_c
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_c
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_c
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x32 aom_dc_left_predictor_32x32_c
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x64 aom_dc_left_predictor_32x64_c
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_c
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_c
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_c
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_64x16 aom_dc_left_predictor_64x16_c
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_64x32 aom_dc_left_predictor_64x32_c
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_64x64 aom_dc_left_predictor_64x64_c
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_c
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_c
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_c
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x16 aom_dc_predictor_16x16_c
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_c
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_c
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x32 aom_dc_predictor_32x32_c
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x64 aom_dc_predictor_32x64_c
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_c
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_c
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x4 aom_dc_predictor_4x4_c
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_64x16 aom_dc_predictor_64x16_c
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_64x32 aom_dc_predictor_64x32_c
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_64x64 aom_dc_predictor_64x64_c
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_c
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_c
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x8 aom_dc_predictor_8x8_c
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_c
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_c
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_c
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x32 aom_dc_top_predictor_32x32_c
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x64 aom_dc_top_predictor_32x64_c
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_c
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_c
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_c
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_64x16 aom_dc_top_predictor_64x16_c
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_64x32 aom_dc_top_predictor_64x32_c
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_64x64 aom_dc_top_predictor_64x64_c
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_c
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_c
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_c
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_comp_avg_pred aom_dist_wtd_comp_avg_pred_c
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x128_avg aom_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x64_avg aom_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x16_avg aom_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x32_avg aom_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x4_avg aom_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x64_avg aom_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x8_avg aom_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x16_avg aom_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x32_avg aom_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x64_avg aom_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x8_avg aom_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x16_avg aom_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x4_avg aom_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x8_avg aom_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x128_avg aom_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x16_avg aom_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x32_avg aom_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x64_avg aom_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x16_avg aom_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x32_avg aom_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x4_avg aom_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x8_avg aom_dist_wtd_sad8x8_avg_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance128x128 aom_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance128x64 aom_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance16x16 aom_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance16x32 aom_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance16x4 aom_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance16x64 aom_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance16x8 aom_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance32x16 aom_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance32x32 aom_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance32x64 aom_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance32x8 aom_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance4x16 aom_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance4x4 aom_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance4x8 aom_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance64x128 aom_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance64x16 aom_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance64x32 aom_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance64x64 aom_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance8x16 aom_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance8x32 aom_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance8x4 aom_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sub_pixel_avg_variance8x8 aom_dist_wtd_sub_pixel_avg_variance8x8_c
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_c
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+#define aom_fdct4x4_lp aom_fdct4x4_lp_c
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+#define aom_fft16x16_float aom_fft16x16_float_c
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+#define aom_fft32x32_float aom_fft32x32_float_c
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+#define aom_fft4x4_float aom_fft4x4_float_c
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+#define aom_fft8x8_float aom_fft8x8_float_c
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+#define aom_get_blk_sse_sum aom_get_blk_sse_sum_c
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_c
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+#define aom_get_var_sse_sum_16x16_dual aom_get_var_sse_sum_16x16_dual_c
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+#define aom_get_var_sse_sum_8x8_quad aom_get_var_sse_sum_8x8_quad_c
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x16 aom_h_predictor_16x16_c
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_c
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_c
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_c
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_c
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_c
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x32 aom_h_predictor_32x32_c
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x64 aom_h_predictor_32x64_c
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_c
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_c
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x4 aom_h_predictor_4x4_c
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_c
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_c
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x32 aom_h_predictor_64x32_c
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x64 aom_h_predictor_64x64_c
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_c
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_c
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_c
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x8 aom_h_predictor_8x8_c
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_16x16 aom_hadamard_16x16_c
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_32x32 aom_hadamard_32x32_c
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_4x4 aom_hadamard_4x4_c
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_8x8 aom_hadamard_8x8_c
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+#define aom_hadamard_lp_16x16 aom_hadamard_lp_16x16_c
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+#define aom_hadamard_lp_8x8 aom_hadamard_lp_8x8_c
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+#define aom_hadamard_lp_8x8_dual aom_hadamard_lp_8x8_dual_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance128x128 aom_highbd_10_masked_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance128x64 aom_highbd_10_masked_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x16 aom_highbd_10_masked_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x32 aom_highbd_10_masked_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x4 aom_highbd_10_masked_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x64 aom_highbd_10_masked_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x8 aom_highbd_10_masked_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance32x16 aom_highbd_10_masked_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance32x32 aom_highbd_10_masked_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance32x64 aom_highbd_10_masked_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance32x8 aom_highbd_10_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance4x16 aom_highbd_10_masked_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance4x4 aom_highbd_10_masked_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance4x8 aom_highbd_10_masked_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance64x128 aom_highbd_10_masked_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance64x16 aom_highbd_10_masked_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance64x32 aom_highbd_10_masked_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance64x64 aom_highbd_10_masked_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance8x16 aom_highbd_10_masked_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance8x32 aom_highbd_10_masked_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance8x4 aom_highbd_10_masked_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance8x8 aom_highbd_10_masked_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_c
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x128 aom_highbd_10_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x64 aom_highbd_10_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x64 aom_highbd_10_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x128 aom_highbd_10_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x16 aom_highbd_10_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance128x128 aom_highbd_10_obmc_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance128x64 aom_highbd_10_obmc_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x16 aom_highbd_10_obmc_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x32 aom_highbd_10_obmc_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x4 aom_highbd_10_obmc_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x64 aom_highbd_10_obmc_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x8 aom_highbd_10_obmc_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance32x16 aom_highbd_10_obmc_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance32x32 aom_highbd_10_obmc_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance32x64 aom_highbd_10_obmc_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance32x8 aom_highbd_10_obmc_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance4x16 aom_highbd_10_obmc_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance4x4 aom_highbd_10_obmc_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance4x8 aom_highbd_10_obmc_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance64x128 aom_highbd_10_obmc_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance64x16 aom_highbd_10_obmc_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance64x32 aom_highbd_10_obmc_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance64x64 aom_highbd_10_obmc_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance8x16 aom_highbd_10_obmc_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance8x32 aom_highbd_10_obmc_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance8x4 aom_highbd_10_obmc_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance8x8 aom_highbd_10_obmc_variance8x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x128 aom_highbd_10_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x64 aom_highbd_10_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x4 aom_highbd_10_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x64 aom_highbd_10_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x8 aom_highbd_10_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x4 aom_highbd_10_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x128 aom_highbd_10_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x16 aom_highbd_10_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x32 aom_highbd_10_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance128x128 aom_highbd_10_sub_pixel_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance128x64 aom_highbd_10_sub_pixel_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x16 aom_highbd_10_sub_pixel_variance16x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x32 aom_highbd_10_sub_pixel_variance16x32_c
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x4 aom_highbd_10_sub_pixel_variance16x4_c
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x64 aom_highbd_10_sub_pixel_variance16x64_c
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x8 aom_highbd_10_sub_pixel_variance16x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x16 aom_highbd_10_sub_pixel_variance32x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x32 aom_highbd_10_sub_pixel_variance32x32_c
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x64 aom_highbd_10_sub_pixel_variance32x64_c
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x8 aom_highbd_10_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x4 aom_highbd_10_sub_pixel_variance4x4_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x128 aom_highbd_10_sub_pixel_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x16 aom_highbd_10_sub_pixel_variance64x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x32 aom_highbd_10_sub_pixel_variance64x32_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x64 aom_highbd_10_sub_pixel_variance64x64_c
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x16 aom_highbd_10_sub_pixel_variance8x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x32 aom_highbd_10_sub_pixel_variance8x32_c
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_c
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x8 aom_highbd_10_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance128x128 aom_highbd_10_variance128x128_c
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance128x64 aom_highbd_10_variance128x64_c
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x16 aom_highbd_10_variance16x16_c
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x32 aom_highbd_10_variance16x32_c
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x64 aom_highbd_10_variance16x64_c
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x8 aom_highbd_10_variance16x8_c
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance32x16 aom_highbd_10_variance32x16_c
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance32x32 aom_highbd_10_variance32x32_c
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance32x64 aom_highbd_10_variance32x64_c
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance32x8 aom_highbd_10_variance32x8_c
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x4 aom_highbd_10_variance4x4_c
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance64x128 aom_highbd_10_variance64x128_c
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance64x16 aom_highbd_10_variance64x16_c
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance64x32 aom_highbd_10_variance64x32_c
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance64x64 aom_highbd_10_variance64x64_c
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x16 aom_highbd_10_variance8x16_c
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x32 aom_highbd_10_variance8x32_c
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x8 aom_highbd_10_variance8x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance128x128 aom_highbd_12_masked_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance128x64 aom_highbd_12_masked_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x16 aom_highbd_12_masked_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x32 aom_highbd_12_masked_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x4 aom_highbd_12_masked_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x64 aom_highbd_12_masked_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x8 aom_highbd_12_masked_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance32x16 aom_highbd_12_masked_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance32x32 aom_highbd_12_masked_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance32x64 aom_highbd_12_masked_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance32x8 aom_highbd_12_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance4x16 aom_highbd_12_masked_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance4x4 aom_highbd_12_masked_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance4x8 aom_highbd_12_masked_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance64x128 aom_highbd_12_masked_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance64x16 aom_highbd_12_masked_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance64x32 aom_highbd_12_masked_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance64x64 aom_highbd_12_masked_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance8x16 aom_highbd_12_masked_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance8x32 aom_highbd_12_masked_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance8x4 aom_highbd_12_masked_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance8x8 aom_highbd_12_masked_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_c
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x128 aom_highbd_12_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x64 aom_highbd_12_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x64 aom_highbd_12_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x128 aom_highbd_12_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x16 aom_highbd_12_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance128x128 aom_highbd_12_obmc_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance128x64 aom_highbd_12_obmc_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x16 aom_highbd_12_obmc_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x32 aom_highbd_12_obmc_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x4 aom_highbd_12_obmc_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x64 aom_highbd_12_obmc_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x8 aom_highbd_12_obmc_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance32x16 aom_highbd_12_obmc_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance32x32 aom_highbd_12_obmc_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance32x64 aom_highbd_12_obmc_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance32x8 aom_highbd_12_obmc_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance4x16 aom_highbd_12_obmc_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance4x4 aom_highbd_12_obmc_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance4x8 aom_highbd_12_obmc_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance64x128 aom_highbd_12_obmc_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance64x16 aom_highbd_12_obmc_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance64x32 aom_highbd_12_obmc_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance64x64 aom_highbd_12_obmc_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance8x16 aom_highbd_12_obmc_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance8x32 aom_highbd_12_obmc_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance8x4 aom_highbd_12_obmc_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance8x8 aom_highbd_12_obmc_variance8x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x128 aom_highbd_12_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x64 aom_highbd_12_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x4 aom_highbd_12_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x64 aom_highbd_12_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x8 aom_highbd_12_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x4 aom_highbd_12_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x128 aom_highbd_12_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x16 aom_highbd_12_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x32 aom_highbd_12_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x128 aom_highbd_12_sub_pixel_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x64 aom_highbd_12_sub_pixel_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_c
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x4 aom_highbd_12_sub_pixel_variance16x4_c
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x64 aom_highbd_12_sub_pixel_variance16x64_c
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_c
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_c
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x8 aom_highbd_12_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x4 aom_highbd_12_sub_pixel_variance4x4_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x128 aom_highbd_12_sub_pixel_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x16 aom_highbd_12_sub_pixel_variance64x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_c
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x32 aom_highbd_12_sub_pixel_variance8x32_c
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_c
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x128 aom_highbd_12_variance128x128_c
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x64 aom_highbd_12_variance128x64_c
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_c
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_c
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x64 aom_highbd_12_variance16x64_c
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_c
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_c
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_c
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_c
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x8 aom_highbd_12_variance32x8_c
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x4 aom_highbd_12_variance4x4_c
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x128 aom_highbd_12_variance64x128_c
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x16 aom_highbd_12_variance64x16_c
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_c
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_c
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_c
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x32 aom_highbd_12_variance8x32_c
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance128x128 aom_highbd_8_masked_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance128x64 aom_highbd_8_masked_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x16 aom_highbd_8_masked_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x32 aom_highbd_8_masked_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x4 aom_highbd_8_masked_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x64 aom_highbd_8_masked_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x8 aom_highbd_8_masked_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance32x16 aom_highbd_8_masked_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance32x32 aom_highbd_8_masked_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance32x64 aom_highbd_8_masked_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance32x8 aom_highbd_8_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance4x16 aom_highbd_8_masked_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance4x4 aom_highbd_8_masked_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance4x8 aom_highbd_8_masked_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance64x128 aom_highbd_8_masked_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance64x16 aom_highbd_8_masked_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance64x32 aom_highbd_8_masked_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance64x64 aom_highbd_8_masked_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance8x16 aom_highbd_8_masked_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance8x32 aom_highbd_8_masked_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance8x4 aom_highbd_8_masked_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance8x8 aom_highbd_8_masked_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_c
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x128 aom_highbd_8_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x64 aom_highbd_8_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x16 aom_highbd_8_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x32 aom_highbd_8_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x4 aom_highbd_8_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x64 aom_highbd_8_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x8 aom_highbd_8_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x16 aom_highbd_8_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x32 aom_highbd_8_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x64 aom_highbd_8_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x8 aom_highbd_8_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x16 aom_highbd_8_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x4 aom_highbd_8_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x8 aom_highbd_8_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x128 aom_highbd_8_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x16 aom_highbd_8_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x32 aom_highbd_8_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x64 aom_highbd_8_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x16 aom_highbd_8_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x32 aom_highbd_8_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x4 aom_highbd_8_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x8 aom_highbd_8_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance128x128 aom_highbd_8_obmc_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance128x64 aom_highbd_8_obmc_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance16x16 aom_highbd_8_obmc_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance16x32 aom_highbd_8_obmc_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance16x4 aom_highbd_8_obmc_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance16x64 aom_highbd_8_obmc_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance16x8 aom_highbd_8_obmc_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance32x16 aom_highbd_8_obmc_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance32x32 aom_highbd_8_obmc_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance32x64 aom_highbd_8_obmc_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance32x8 aom_highbd_8_obmc_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance4x16 aom_highbd_8_obmc_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance4x4 aom_highbd_8_obmc_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance4x8 aom_highbd_8_obmc_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance64x128 aom_highbd_8_obmc_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance64x16 aom_highbd_8_obmc_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance64x32 aom_highbd_8_obmc_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance64x64 aom_highbd_8_obmc_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance8x16 aom_highbd_8_obmc_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance8x32 aom_highbd_8_obmc_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance8x4 aom_highbd_8_obmc_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_variance8x8 aom_highbd_8_obmc_variance8x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x128 aom_highbd_8_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x64 aom_highbd_8_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x4 aom_highbd_8_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x64 aom_highbd_8_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x8 aom_highbd_8_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x4 aom_highbd_8_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x128 aom_highbd_8_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x16 aom_highbd_8_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x32 aom_highbd_8_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x128 aom_highbd_8_sub_pixel_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x64 aom_highbd_8_sub_pixel_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_c
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x4 aom_highbd_8_sub_pixel_variance16x4_c
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x64 aom_highbd_8_sub_pixel_variance16x64_c
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_c
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_c
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x8 aom_highbd_8_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x4 aom_highbd_8_sub_pixel_variance4x4_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x128 aom_highbd_8_sub_pixel_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x16 aom_highbd_8_sub_pixel_variance64x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_c
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x32 aom_highbd_8_sub_pixel_variance8x32_c
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_c
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x128 aom_highbd_8_variance128x128_c
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x64 aom_highbd_8_variance128x64_c
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_c
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_c
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x64 aom_highbd_8_variance16x64_c
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_c
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_c
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_c
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_c
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x8 aom_highbd_8_variance32x8_c
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x4 aom_highbd_8_variance4x4_c
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x128 aom_highbd_8_variance64x128_c
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x16 aom_highbd_8_variance64x16_c
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_c
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_c
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_c
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x32 aom_highbd_8_variance8x32_c
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_c
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+#define aom_highbd_avg_4x4 aom_highbd_avg_4x4_c
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+#define aom_highbd_avg_8x8 aom_highbd_avg_8x8_c
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+#define aom_highbd_blend_a64_d16_mask aom_highbd_blend_a64_d16_mask_c
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+#define aom_highbd_blend_a64_hmask aom_highbd_blend_a64_hmask_c
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+#define aom_highbd_blend_a64_mask aom_highbd_blend_a64_mask_c
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+#define aom_highbd_blend_a64_vmask aom_highbd_blend_a64_vmask_c
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+#define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+#define aom_highbd_comp_mask_pred aom_highbd_comp_mask_pred_c
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+#define aom_highbd_convolve8_horiz aom_highbd_convolve8_horiz_c
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+#define aom_highbd_convolve8_vert aom_highbd_convolve8_vert_c
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+#define aom_highbd_convolve_copy aom_highbd_convolve_copy_c
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_c
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_c
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x4 aom_highbd_dc_128_predictor_16x4_c
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x64 aom_highbd_dc_128_predictor_16x64_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_c
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x64 aom_highbd_dc_128_predictor_32x64_c
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x8 aom_highbd_dc_128_predictor_32x8_c
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x16 aom_highbd_dc_128_predictor_4x16_c
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_c
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_c
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x16 aom_highbd_dc_128_predictor_64x16_c
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x32 aom_highbd_dc_128_predictor_64x32_c
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x64 aom_highbd_dc_128_predictor_64x64_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_c
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x32 aom_highbd_dc_128_predictor_8x32_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_c
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_c
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_c
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_c
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x4 aom_highbd_dc_left_predictor_16x4_c
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x64 aom_highbd_dc_left_predictor_16x64_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_c
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x64 aom_highbd_dc_left_predictor_32x64_c
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x8 aom_highbd_dc_left_predictor_32x8_c
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x16 aom_highbd_dc_left_predictor_4x16_c
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_c
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_c
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x16 aom_highbd_dc_left_predictor_64x16_c
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x32 aom_highbd_dc_left_predictor_64x32_c
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x64 aom_highbd_dc_left_predictor_64x64_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_c
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x32 aom_highbd_dc_left_predictor_8x32_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_c
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_c
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_c
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_c
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x4 aom_highbd_dc_predictor_16x4_c
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x64 aom_highbd_dc_predictor_16x64_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_c
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x64 aom_highbd_dc_predictor_32x64_c
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x8 aom_highbd_dc_predictor_32x8_c
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x16 aom_highbd_dc_predictor_4x16_c
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_c
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_c
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x16 aom_highbd_dc_predictor_64x16_c
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x32 aom_highbd_dc_predictor_64x32_c
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x64 aom_highbd_dc_predictor_64x64_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_c
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x32 aom_highbd_dc_predictor_8x32_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_c
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_c
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_c
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_c
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x4 aom_highbd_dc_top_predictor_16x4_c
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x64 aom_highbd_dc_top_predictor_16x64_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_c
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x64 aom_highbd_dc_top_predictor_32x64_c
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x8 aom_highbd_dc_top_predictor_32x8_c
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x16 aom_highbd_dc_top_predictor_4x16_c
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_c
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_c
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x16 aom_highbd_dc_top_predictor_64x16_c
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x32 aom_highbd_dc_top_predictor_64x32_c
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x64 aom_highbd_dc_top_predictor_64x64_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_c
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x32 aom_highbd_dc_top_predictor_8x32_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_c
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_c
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_highbd_dist_wtd_comp_avg_pred aom_highbd_dist_wtd_comp_avg_pred_c
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_c
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_c
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x4 aom_highbd_h_predictor_16x4_c
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x64 aom_highbd_h_predictor_16x64_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_c
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_c
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_c
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x64 aom_highbd_h_predictor_32x64_c
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x8 aom_highbd_h_predictor_32x8_c
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x16 aom_highbd_h_predictor_4x16_c
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_c
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_c
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x16 aom_highbd_h_predictor_64x16_c
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x32 aom_highbd_h_predictor_64x32_c
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x64 aom_highbd_h_predictor_64x64_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_c
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x32 aom_highbd_h_predictor_8x32_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_c
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_c
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_highbd_hadamard_16x16 aom_highbd_hadamard_16x16_c
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_highbd_hadamard_32x32 aom_highbd_hadamard_32x32_c
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_highbd_hadamard_8x8 aom_highbd_hadamard_8x8_c
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_14 aom_highbd_lpf_horizontal_14_c
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+#define aom_highbd_lpf_horizontal_14_dual aom_highbd_lpf_horizontal_14_dual_c
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_c
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_4_dual aom_highbd_lpf_horizontal_4_dual_c
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_6 aom_highbd_lpf_horizontal_6_c
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_6_dual aom_highbd_lpf_horizontal_6_dual_c
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_8 aom_highbd_lpf_horizontal_8_c
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_8_dual aom_highbd_lpf_horizontal_8_dual_c
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_14 aom_highbd_lpf_vertical_14_c
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_14_dual aom_highbd_lpf_vertical_14_dual_c
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_4 aom_highbd_lpf_vertical_4_c
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_4_dual aom_highbd_lpf_vertical_4_dual_c
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_6 aom_highbd_lpf_vertical_6_c
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_6_dual aom_highbd_lpf_vertical_6_dual_c
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_8 aom_highbd_lpf_vertical_8_c
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_8_dual aom_highbd_lpf_vertical_8_dual_c
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad128x128 aom_highbd_masked_sad128x128_c
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad128x64 aom_highbd_masked_sad128x64_c
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad16x16 aom_highbd_masked_sad16x16_c
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad16x32 aom_highbd_masked_sad16x32_c
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad16x4 aom_highbd_masked_sad16x4_c
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad16x64 aom_highbd_masked_sad16x64_c
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad16x8 aom_highbd_masked_sad16x8_c
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad32x16 aom_highbd_masked_sad32x16_c
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad32x32 aom_highbd_masked_sad32x32_c
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad32x64 aom_highbd_masked_sad32x64_c
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad32x8 aom_highbd_masked_sad32x8_c
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad4x16 aom_highbd_masked_sad4x16_c
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad4x4 aom_highbd_masked_sad4x4_c
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad4x8 aom_highbd_masked_sad4x8_c
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad64x128 aom_highbd_masked_sad64x128_c
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad64x16 aom_highbd_masked_sad64x16_c
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad64x32 aom_highbd_masked_sad64x32_c
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad64x64 aom_highbd_masked_sad64x64_c
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad8x16 aom_highbd_masked_sad8x16_c
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad8x32 aom_highbd_masked_sad8x32_c
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad8x4 aom_highbd_masked_sad8x4_c
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad8x8 aom_highbd_masked_sad8x8_c
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad128x128 aom_highbd_obmc_sad128x128_c
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad128x64 aom_highbd_obmc_sad128x64_c
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad16x16 aom_highbd_obmc_sad16x16_c
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad16x32 aom_highbd_obmc_sad16x32_c
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad16x4 aom_highbd_obmc_sad16x4_c
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad16x64 aom_highbd_obmc_sad16x64_c
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad16x8 aom_highbd_obmc_sad16x8_c
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad32x16 aom_highbd_obmc_sad32x16_c
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad32x32 aom_highbd_obmc_sad32x32_c
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad32x64 aom_highbd_obmc_sad32x64_c
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad32x8 aom_highbd_obmc_sad32x8_c
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad4x16 aom_highbd_obmc_sad4x16_c
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad4x4 aom_highbd_obmc_sad4x4_c
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad4x8 aom_highbd_obmc_sad4x8_c
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad64x128 aom_highbd_obmc_sad64x128_c
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad64x16 aom_highbd_obmc_sad64x16_c
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad64x32 aom_highbd_obmc_sad64x32_c
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad64x64 aom_highbd_obmc_sad64x64_c
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad8x16 aom_highbd_obmc_sad8x16_c
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad8x32 aom_highbd_obmc_sad8x32_c
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad8x4 aom_highbd_obmc_sad8x4_c
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad8x8 aom_highbd_obmc_sad8x8_c
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b aom_highbd_quantize_b_c
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_32x32 aom_highbd_quantize_b_32x32_c
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_32x32_adaptive aom_highbd_quantize_b_32x32_adaptive_c
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_64x64 aom_highbd_quantize_b_64x64_c
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_64x64_adaptive aom_highbd_quantize_b_64x64_adaptive_c
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_adaptive aom_highbd_quantize_b_adaptive_c
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad128x128 aom_highbd_sad128x128_c
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad128x128_avg aom_highbd_sad128x128_avg_c
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad128x128x3d aom_highbd_sad128x128x3d_c
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad128x128x4d aom_highbd_sad128x128x4d_c
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad128x64 aom_highbd_sad128x64_c
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad128x64_avg aom_highbd_sad128x64_avg_c
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad128x64x3d aom_highbd_sad128x64x3d_c
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad128x64x4d aom_highbd_sad128x64x4d_c
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad16x16 aom_highbd_sad16x16_c
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad16x16_avg aom_highbd_sad16x16_avg_c
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x16x3d aom_highbd_sad16x16x3d_c
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x16x4d aom_highbd_sad16x16x4d_c
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad16x32 aom_highbd_sad16x32_c
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad16x32_avg aom_highbd_sad16x32_avg_c
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x32x3d aom_highbd_sad16x32x3d_c
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x32x4d aom_highbd_sad16x32x4d_c
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad16x4 aom_highbd_sad16x4_c
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad16x4_avg aom_highbd_sad16x4_avg_c
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x4x3d aom_highbd_sad16x4x3d_c
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x4x4d aom_highbd_sad16x4x4d_c
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad16x64 aom_highbd_sad16x64_c
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad16x64_avg aom_highbd_sad16x64_avg_c
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x64x3d aom_highbd_sad16x64x3d_c
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x64x4d aom_highbd_sad16x64x4d_c
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad16x8 aom_highbd_sad16x8_c
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad16x8_avg aom_highbd_sad16x8_avg_c
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x8x3d aom_highbd_sad16x8x3d_c
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad16x8x4d aom_highbd_sad16x8x4d_c
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad32x16 aom_highbd_sad32x16_c
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad32x16_avg aom_highbd_sad32x16_avg_c
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x16x3d aom_highbd_sad32x16x3d_c
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x16x4d aom_highbd_sad32x16x4d_c
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad32x32 aom_highbd_sad32x32_c
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad32x32_avg aom_highbd_sad32x32_avg_c
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x32x3d aom_highbd_sad32x32x3d_c
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x32x4d aom_highbd_sad32x32x4d_c
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad32x64 aom_highbd_sad32x64_c
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad32x64_avg aom_highbd_sad32x64_avg_c
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x64x3d aom_highbd_sad32x64x3d_c
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x64x4d aom_highbd_sad32x64x4d_c
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad32x8 aom_highbd_sad32x8_c
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad32x8_avg aom_highbd_sad32x8_avg_c
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x8x3d aom_highbd_sad32x8x3d_c
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad32x8x4d aom_highbd_sad32x8x4d_c
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x16 aom_highbd_sad4x16_c
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x16_avg aom_highbd_sad4x16_avg_c
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x3d aom_highbd_sad4x16x3d_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x4d aom_highbd_sad4x16x4d_c
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x4 aom_highbd_sad4x4_c
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x4_avg aom_highbd_sad4x4_avg_c
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x3d aom_highbd_sad4x4x3d_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x4d aom_highbd_sad4x4x4d_c
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x8 aom_highbd_sad4x8_c
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x8_avg aom_highbd_sad4x8_avg_c
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x3d aom_highbd_sad4x8x3d_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x4d aom_highbd_sad4x8x4d_c
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad64x128 aom_highbd_sad64x128_c
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad64x128_avg aom_highbd_sad64x128_avg_c
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x128x3d aom_highbd_sad64x128x3d_c
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x128x4d aom_highbd_sad64x128x4d_c
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad64x16 aom_highbd_sad64x16_c
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad64x16_avg aom_highbd_sad64x16_avg_c
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x16x3d aom_highbd_sad64x16x3d_c
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x16x4d aom_highbd_sad64x16x4d_c
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad64x32 aom_highbd_sad64x32_c
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad64x32_avg aom_highbd_sad64x32_avg_c
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x32x3d aom_highbd_sad64x32x3d_c
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x32x4d aom_highbd_sad64x32x4d_c
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad64x64 aom_highbd_sad64x64_c
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad64x64_avg aom_highbd_sad64x64_avg_c
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x64x3d aom_highbd_sad64x64x3d_c
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad64x64x4d aom_highbd_sad64x64x4d_c
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x16 aom_highbd_sad8x16_c
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x16_avg aom_highbd_sad8x16_avg_c
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x3d aom_highbd_sad8x16x3d_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x4d aom_highbd_sad8x16x4d_c
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x32 aom_highbd_sad8x32_c
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x32_avg aom_highbd_sad8x32_avg_c
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x3d aom_highbd_sad8x32x3d_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x4d aom_highbd_sad8x32x4d_c
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x4 aom_highbd_sad8x4_c
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x4_avg aom_highbd_sad8x4_avg_c
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x3d aom_highbd_sad8x4x3d_c
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x4d aom_highbd_sad8x4x4d_c
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x8 aom_highbd_sad8x8_c
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x8_avg aom_highbd_sad8x8_avg_c
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x3d aom_highbd_sad8x8x3d_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x4d aom_highbd_sad8x8x4d_c
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_128x128 aom_highbd_sad_skip_128x128_c
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_128x128x4d aom_highbd_sad_skip_128x128x4d_c
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_128x64 aom_highbd_sad_skip_128x64_c
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_128x64x4d aom_highbd_sad_skip_128x64x4d_c
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x16 aom_highbd_sad_skip_16x16_c
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x16x4d aom_highbd_sad_skip_16x16x4d_c
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x32 aom_highbd_sad_skip_16x32_c
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x32x4d aom_highbd_sad_skip_16x32x4d_c
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x4 aom_highbd_sad_skip_16x4_c
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x4x4d aom_highbd_sad_skip_16x4x4d_c
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x64 aom_highbd_sad_skip_16x64_c
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x64x4d aom_highbd_sad_skip_16x64x4d_c
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x8 aom_highbd_sad_skip_16x8_c
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x8x4d aom_highbd_sad_skip_16x8x4d_c
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_32x16 aom_highbd_sad_skip_32x16_c
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_32x16x4d aom_highbd_sad_skip_32x16x4d_c
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_32x32 aom_highbd_sad_skip_32x32_c
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_32x32x4d aom_highbd_sad_skip_32x32x4d_c
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_32x64 aom_highbd_sad_skip_32x64_c
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_32x64x4d aom_highbd_sad_skip_32x64x4d_c
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_32x8 aom_highbd_sad_skip_32x8_c
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_32x8x4d aom_highbd_sad_skip_32x8x4d_c
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x16 aom_highbd_sad_skip_4x16_c
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x16x4d aom_highbd_sad_skip_4x16x4d_c
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x4 aom_highbd_sad_skip_4x4_c
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x4x4d aom_highbd_sad_skip_4x4x4d_c
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x8 aom_highbd_sad_skip_4x8_c
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x8x4d aom_highbd_sad_skip_4x8x4d_c
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_64x128 aom_highbd_sad_skip_64x128_c
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_64x128x4d aom_highbd_sad_skip_64x128x4d_c
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_64x16 aom_highbd_sad_skip_64x16_c
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_64x16x4d aom_highbd_sad_skip_64x16x4d_c
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_64x32 aom_highbd_sad_skip_64x32_c
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_64x32x4d aom_highbd_sad_skip_64x32x4d_c
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_64x64 aom_highbd_sad_skip_64x64_c
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_64x64x4d aom_highbd_sad_skip_64x64x4d_c
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x16 aom_highbd_sad_skip_8x16_c
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x16x4d aom_highbd_sad_skip_8x16x4d_c
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x32 aom_highbd_sad_skip_8x32_c
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x32x4d aom_highbd_sad_skip_8x32x4d_c
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x4 aom_highbd_sad_skip_8x4_c
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x4x4d aom_highbd_sad_skip_8x4x4d_c
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x8 aom_highbd_sad_skip_8x8_c
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x8x4d aom_highbd_sad_skip_8x8x4d_c
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+#define aom_highbd_sse aom_highbd_sse_c
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define aom_highbd_subtract_block aom_highbd_subtract_block_c
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_c
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_c
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_c
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_c
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_c
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_c
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_c
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_c
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_c
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_c
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+#define aom_ifft16x16_float aom_ifft16x16_float_c
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+#define aom_ifft32x32_float aom_ifft32x32_float_c
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+#define aom_ifft4x4_float aom_ifft4x4_float_c
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+#define aom_ifft8x8_float aom_ifft8x8_float_c
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+#define aom_int_pro_col aom_int_pro_col_c
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+#define aom_int_pro_row aom_int_pro_row_c
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+#define aom_lowbd_blend_a64_d16_mask aom_lowbd_blend_a64_d16_mask_c
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_14 aom_lpf_horizontal_14_c
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_c
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_4 aom_lpf_horizontal_4_c
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_c
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_6 aom_lpf_horizontal_6_c
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_c
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_8 aom_lpf_horizontal_8_c
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_c
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_14 aom_lpf_vertical_14_c
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_c
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_4 aom_lpf_vertical_4_c
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_c
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_6 aom_lpf_vertical_6_c
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_c
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_8 aom_lpf_vertical_8_c
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_c
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad128x128 aom_masked_sad128x128_c
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad128x128x4d aom_masked_sad128x128x4d_c
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad128x64 aom_masked_sad128x64_c
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad128x64x4d aom_masked_sad128x64x4d_c
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad16x16 aom_masked_sad16x16_c
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad16x16x4d aom_masked_sad16x16x4d_c
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad16x32 aom_masked_sad16x32_c
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad16x32x4d aom_masked_sad16x32x4d_c
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad16x4 aom_masked_sad16x4_c
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad16x4x4d aom_masked_sad16x4x4d_c
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad16x64 aom_masked_sad16x64_c
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad16x64x4d aom_masked_sad16x64x4d_c
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad16x8 aom_masked_sad16x8_c
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad16x8x4d aom_masked_sad16x8x4d_c
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad32x16 aom_masked_sad32x16_c
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad32x16x4d aom_masked_sad32x16x4d_c
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad32x32 aom_masked_sad32x32_c
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad32x32x4d aom_masked_sad32x32x4d_c
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad32x64 aom_masked_sad32x64_c
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad32x64x4d aom_masked_sad32x64x4d_c
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad32x8 aom_masked_sad32x8_c
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad32x8x4d aom_masked_sad32x8x4d_c
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad4x16 aom_masked_sad4x16_c
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad4x16x4d aom_masked_sad4x16x4d_c
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad4x4 aom_masked_sad4x4_c
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad4x4x4d aom_masked_sad4x4x4d_c
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad4x8 aom_masked_sad4x8_c
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad4x8x4d aom_masked_sad4x8x4d_c
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad64x128 aom_masked_sad64x128_c
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad64x128x4d aom_masked_sad64x128x4d_c
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad64x16 aom_masked_sad64x16_c
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad64x16x4d aom_masked_sad64x16x4d_c
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad64x32 aom_masked_sad64x32_c
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad64x32x4d aom_masked_sad64x32x4d_c
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad64x64 aom_masked_sad64x64_c
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad64x64x4d aom_masked_sad64x64x4d_c
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad8x16 aom_masked_sad8x16_c
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad8x16x4d aom_masked_sad8x16x4d_c
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad8x32 aom_masked_sad8x32_c
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad8x32x4d aom_masked_sad8x32x4d_c
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad8x4 aom_masked_sad8x4_c
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad8x4x4d aom_masked_sad8x4x4d_c
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad8x8 aom_masked_sad8x8_c
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+#define aom_masked_sad8x8x4d aom_masked_sad8x8x4d_c
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance128x128 aom_masked_sub_pixel_variance128x128_c
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance128x64 aom_masked_sub_pixel_variance128x64_c
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance16x16 aom_masked_sub_pixel_variance16x16_c
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance16x32 aom_masked_sub_pixel_variance16x32_c
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance16x4 aom_masked_sub_pixel_variance16x4_c
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance16x64 aom_masked_sub_pixel_variance16x64_c
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance16x8 aom_masked_sub_pixel_variance16x8_c
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance32x16 aom_masked_sub_pixel_variance32x16_c
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance32x32 aom_masked_sub_pixel_variance32x32_c
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance32x64 aom_masked_sub_pixel_variance32x64_c
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance32x8 aom_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance4x16 aom_masked_sub_pixel_variance4x16_c
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance4x4 aom_masked_sub_pixel_variance4x4_c
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance4x8 aom_masked_sub_pixel_variance4x8_c
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance64x128 aom_masked_sub_pixel_variance64x128_c
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance64x16 aom_masked_sub_pixel_variance64x16_c
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance64x32 aom_masked_sub_pixel_variance64x32_c
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance64x64 aom_masked_sub_pixel_variance64x64_c
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance8x16 aom_masked_sub_pixel_variance8x16_c
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance8x32 aom_masked_sub_pixel_variance8x32_c
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance8x4 aom_masked_sub_pixel_variance8x4_c
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance8x8 aom_masked_sub_pixel_variance8x8_c
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_minmax_8x8 aom_minmax_8x8_c
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x16 aom_mse16x16_c
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_c
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_c
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_c
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+#define aom_mse_16xh_16bit aom_mse_16xh_16bit_c
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+#define aom_mse_wxh_16bit aom_mse_wxh_16bit_c
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+#define aom_mse_wxh_16bit_highbd aom_mse_wxh_16bit_highbd_c
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad128x128 aom_obmc_sad128x128_c
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad128x64 aom_obmc_sad128x64_c
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x16 aom_obmc_sad16x16_c
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x32 aom_obmc_sad16x32_c
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x4 aom_obmc_sad16x4_c
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x64 aom_obmc_sad16x64_c
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x8 aom_obmc_sad16x8_c
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x16 aom_obmc_sad32x16_c
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x32 aom_obmc_sad32x32_c
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x64 aom_obmc_sad32x64_c
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x8 aom_obmc_sad32x8_c
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x16 aom_obmc_sad4x16_c
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x4 aom_obmc_sad4x4_c
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x8 aom_obmc_sad4x8_c
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x128 aom_obmc_sad64x128_c
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x16 aom_obmc_sad64x16_c
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x32 aom_obmc_sad64x32_c
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x64 aom_obmc_sad64x64_c
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x16 aom_obmc_sad8x16_c
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x32 aom_obmc_sad8x32_c
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x4 aom_obmc_sad8x4_c
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x8 aom_obmc_sad8x8_c
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance128x128 aom_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance128x64 aom_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x16 aom_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x32 aom_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x4 aom_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x64 aom_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x8 aom_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x16 aom_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x32 aom_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x64 aom_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x8 aom_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x16 aom_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x4 aom_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x8 aom_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x128 aom_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x16 aom_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x32 aom_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x64 aom_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x16 aom_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x32 aom_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x4 aom_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x8 aom_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance128x128 aom_obmc_variance128x128_c
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance128x64 aom_obmc_variance128x64_c
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x16 aom_obmc_variance16x16_c
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x32 aom_obmc_variance16x32_c
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x4 aom_obmc_variance16x4_c
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x64 aom_obmc_variance16x64_c
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x8 aom_obmc_variance16x8_c
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x16 aom_obmc_variance32x16_c
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x32 aom_obmc_variance32x32_c
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x64 aom_obmc_variance32x64_c
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x8 aom_obmc_variance32x8_c
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x16 aom_obmc_variance4x16_c
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x4 aom_obmc_variance4x4_c
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x8 aom_obmc_variance4x8_c
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x128 aom_obmc_variance64x128_c
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x16 aom_obmc_variance64x16_c
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x32 aom_obmc_variance64x32_c
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x64 aom_obmc_variance64x64_c
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x16 aom_obmc_variance8x16_c
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x32 aom_obmc_variance8x32_c
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x4 aom_obmc_variance8x4_c
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x8 aom_obmc_variance8x8_c
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_c
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_c
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_c
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_c
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_c
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_c
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b aom_quantize_b_c
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32 aom_quantize_b_32x32_c
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32_adaptive aom_quantize_b_32x32_adaptive_c
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64 aom_quantize_b_64x64_c
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64_adaptive aom_quantize_b_64x64_adaptive_c
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_adaptive aom_quantize_b_adaptive_c
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad128x128 aom_sad128x128_c
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad128x128_avg aom_sad128x128_avg_c
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad128x128x3d aom_sad128x128x3d_c
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad128x128x4d aom_sad128x128x4d_c
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad128x64 aom_sad128x64_c
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad128x64_avg aom_sad128x64_avg_c
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad128x64x3d aom_sad128x64x3d_c
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad128x64x4d aom_sad128x64x4d_c
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x16 aom_sad16x16_c
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_c
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x16x3d aom_sad16x16x3d_c
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x16x4d aom_sad16x16x4d_c
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_c
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_c
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x32x3d aom_sad16x32x3d_c
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x32x4d aom_sad16x32x4d_c
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x4 aom_sad16x4_c
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x4_avg aom_sad16x4_avg_c
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x3d aom_sad16x4x3d_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x4d aom_sad16x4x4d_c
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x64 aom_sad16x64_c
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x64_avg aom_sad16x64_avg_c
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x64x3d aom_sad16x64x3d_c
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x64x4d aom_sad16x64x4d_c
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x8 aom_sad16x8_c
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_c
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x8x3d aom_sad16x8x3d_c
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x8x4d aom_sad16x8x4d_c
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x16 aom_sad32x16_c
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x16_avg aom_sad32x16_avg_c
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x16x3d aom_sad32x16x3d_c
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x16x4d aom_sad32x16x4d_c
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x32 aom_sad32x32_c
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x32_avg aom_sad32x32_avg_c
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x32x3d aom_sad32x32x3d_c
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x32x4d aom_sad32x32x4d_c
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x64 aom_sad32x64_c
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x64_avg aom_sad32x64_avg_c
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x64x3d aom_sad32x64x3d_c
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x64x4d aom_sad32x64x4d_c
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x8 aom_sad32x8_c
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x8_avg aom_sad32x8_avg_c
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x8x3d aom_sad32x8x3d_c
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad32x8x4d aom_sad32x8x4d_c
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x16 aom_sad4x16_c
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x16_avg aom_sad4x16_avg_c
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x3d aom_sad4x16x3d_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x4d aom_sad4x16x4d_c
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x4 aom_sad4x4_c
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_c
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x3d aom_sad4x4x3d_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x4d aom_sad4x4x4d_c
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_c
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_c
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x3d aom_sad4x8x3d_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x4d aom_sad4x8x4d_c
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x128 aom_sad64x128_c
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x128_avg aom_sad64x128_avg_c
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x128x3d aom_sad64x128x3d_c
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x128x4d aom_sad64x128x4d_c
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x16 aom_sad64x16_c
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x16_avg aom_sad64x16_avg_c
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x16x3d aom_sad64x16x3d_c
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x16x4d aom_sad64x16x4d_c
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x32 aom_sad64x32_c
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x32_avg aom_sad64x32_avg_c
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x32x3d aom_sad64x32x3d_c
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x32x4d aom_sad64x32x4d_c
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x64 aom_sad64x64_c
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x64_avg aom_sad64x64_avg_c
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x64x3d aom_sad64x64x3d_c
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad64x64x4d aom_sad64x64x4d_c
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x16 aom_sad8x16_c
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_c
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x3d aom_sad8x16x3d_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x4d aom_sad8x16x4d_c
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x32 aom_sad8x32_c
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x32_avg aom_sad8x32_avg_c
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x3d aom_sad8x32x3d_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x4d aom_sad8x32x4d_c
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_c
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_c
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x3d aom_sad8x4x3d_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x4d aom_sad8x4x4d_c
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x8 aom_sad8x8_c
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_c
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x3d aom_sad8x8x3d_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x4d aom_sad8x8x4d_c
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_128x128 aom_sad_skip_128x128_c
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_128x128x4d aom_sad_skip_128x128x4d_c
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_128x64 aom_sad_skip_128x64_c
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_128x64x4d aom_sad_skip_128x64x4d_c
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x16 aom_sad_skip_16x16_c
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x16x4d aom_sad_skip_16x16x4d_c
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x32 aom_sad_skip_16x32_c
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x32x4d aom_sad_skip_16x32x4d_c
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x4 aom_sad_skip_16x4_c
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x64 aom_sad_skip_16x64_c
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x64x4d aom_sad_skip_16x64x4d_c
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x8 aom_sad_skip_16x8_c
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x8x4d aom_sad_skip_16x8x4d_c
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x16 aom_sad_skip_32x16_c
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_32x16x4d aom_sad_skip_32x16x4d_c
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x32 aom_sad_skip_32x32_c
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_32x32x4d aom_sad_skip_32x32x4d_c
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x64 aom_sad_skip_32x64_c
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_32x64x4d aom_sad_skip_32x64x4d_c
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x8 aom_sad_skip_32x8_c
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_32x8x4d aom_sad_skip_32x8x4d_c
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x16 aom_sad_skip_4x16_c
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x16x4d aom_sad_skip_4x16x4d_c
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x4 aom_sad_skip_4x4_c
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x4x4d aom_sad_skip_4x4x4d_c
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x8 aom_sad_skip_4x8_c
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x8x4d aom_sad_skip_4x8x4d_c
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x128 aom_sad_skip_64x128_c
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_64x128x4d aom_sad_skip_64x128x4d_c
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x16 aom_sad_skip_64x16_c
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_64x16x4d aom_sad_skip_64x16x4d_c
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x32 aom_sad_skip_64x32_c
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_64x32x4d aom_sad_skip_64x32x4d_c
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x64 aom_sad_skip_64x64_c
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_64x64x4d aom_sad_skip_64x64x4d_c
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x16 aom_sad_skip_8x16_c
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x16x4d aom_sad_skip_8x16x4d_c
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x32 aom_sad_skip_8x32_c
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x32x4d aom_sad_skip_8x32x4d_c
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x4 aom_sad_skip_8x4_c
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x4x4d aom_sad_skip_8x4x4d_c
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x8 aom_sad_skip_8x8_c
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x8x4d aom_sad_skip_8x8x4d_c
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+#define aom_satd aom_satd_c
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+#define aom_satd_lp aom_satd_lp_c
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_scaled_2d aom_scaled_2d_c
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_c
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_c
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_c
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_c
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_c
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_c
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x16 aom_smooth_predictor_16x16_c
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_c
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_c
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_32x16 aom_smooth_predictor_32x16_c
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_32x32 aom_smooth_predictor_32x32_c
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_c
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_c
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_c
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_4x4 aom_smooth_predictor_4x4_c
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_c
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_64x32 aom_smooth_predictor_64x32_c
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_64x64 aom_smooth_predictor_64x64_c
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_c
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x4 aom_smooth_predictor_8x4_c
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x8 aom_smooth_predictor_8x8_c
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_c
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_c
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_c
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_c
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_c
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_c
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+#define aom_sse aom_sse_c
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_c
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance128x128 aom_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance128x64 aom_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x16 aom_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x32 aom_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x4 aom_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x64 aom_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x8 aom_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x16 aom_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x32 aom_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x64 aom_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x8 aom_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x16 aom_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x4 aom_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x8 aom_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x128 aom_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x16 aom_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x32 aom_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x64 aom_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x16 aom_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x32 aom_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x4 aom_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x8 aom_sub_pixel_avg_variance8x8_c
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance128x128 aom_sub_pixel_variance128x128_c
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance128x64 aom_sub_pixel_variance128x64_c
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x16 aom_sub_pixel_variance16x16_c
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x32 aom_sub_pixel_variance16x32_c
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x4 aom_sub_pixel_variance16x4_c
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x64 aom_sub_pixel_variance16x64_c
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x8 aom_sub_pixel_variance16x8_c
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x16 aom_sub_pixel_variance32x16_c
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x32 aom_sub_pixel_variance32x32_c
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x64 aom_sub_pixel_variance32x64_c
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x8 aom_sub_pixel_variance32x8_c
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x16 aom_sub_pixel_variance4x16_c
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x4 aom_sub_pixel_variance4x4_c
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x8 aom_sub_pixel_variance4x8_c
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x128 aom_sub_pixel_variance64x128_c
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x16 aom_sub_pixel_variance64x16_c
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x32 aom_sub_pixel_variance64x32_c
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x64 aom_sub_pixel_variance64x64_c
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x16 aom_sub_pixel_variance8x16_c
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x32 aom_sub_pixel_variance8x32_c
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x4 aom_sub_pixel_variance8x4_c
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x8 aom_sub_pixel_variance8x8_c
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define aom_subtract_block aom_subtract_block_c
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+#define aom_sum_squares_2d_i16 aom_sum_squares_2d_i16_c
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+#define aom_sum_squares_i16 aom_sum_squares_i16_c
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+#define aom_sum_sse_2d_i16 aom_sum_sse_2d_i16_c
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x16 aom_v_predictor_16x16_c
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_c
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_c
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_c
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_c
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x16 aom_v_predictor_32x16_c
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x32 aom_v_predictor_32x32_c
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x64 aom_v_predictor_32x64_c
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_c
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_c
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x4 aom_v_predictor_4x4_c
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_c
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_64x16 aom_v_predictor_64x16_c
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_64x32 aom_v_predictor_64x32_c
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_64x64 aom_v_predictor_64x64_c
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_c
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_c
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_c
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x8 aom_v_predictor_8x8_c
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+#define aom_var_2d_u16 aom_var_2d_u16_c
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+#define aom_var_2d_u8 aom_var_2d_u8_c
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance128x128 aom_variance128x128_c
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance128x64 aom_variance128x64_c
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x16 aom_variance16x16_c
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x32 aom_variance16x32_c
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x4 aom_variance16x4_c
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x64 aom_variance16x64_c
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x8 aom_variance16x8_c
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x16 aom_variance32x16_c
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x32 aom_variance32x32_c
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x64 aom_variance32x64_c
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x8 aom_variance32x8_c
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x16 aom_variance4x16_c
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x4 aom_variance4x4_c
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x8 aom_variance4x8_c
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance64x128 aom_variance64x128_c
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance64x16 aom_variance64x16_c
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance64x32 aom_variance64x32_c
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance64x64 aom_variance64x64_c
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x16 aom_variance8x16_c
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x32 aom_variance8x32_c
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x4 aom_variance8x4_c
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x8 aom_variance8x8_c
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+#define aom_vector_var aom_vector_var_c
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+#define av1_compute_cross_correlation av1_compute_cross_correlation_c
+
+void aom_dsp_rtcd(void);
+
+#include "config/aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/generic/config/aom_scale_rtcd.h b/media/libaom/config/generic/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..733b2d9ea1
--- /dev/null
+++ b/media/libaom/config/generic/config/aom_scale_rtcd.h
@@ -0,0 +1,100 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#include "config/aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/generic/config/av1_rtcd.h b/media/libaom/config/generic/config/av1_rtcd.h
new file mode 100644
index 0000000000..a6bddea376
--- /dev/null
+++ b/media/libaom/config/generic/config/av1_rtcd.h
@@ -0,0 +1,641 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_c
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+#define aom_dist_wtd_comp_avg_upsampled_pred aom_dist_wtd_comp_avg_upsampled_pred_c
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_c
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+#define aom_highbd_dist_wtd_comp_avg_upsampled_pred aom_highbd_dist_wtd_comp_avg_upsampled_pred_c
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_upsampled_pred aom_highbd_upsampled_pred_c
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+#define aom_quantize_b_helper aom_quantize_b_helper_c
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+#define aom_upsampled_pred aom_upsampled_pred_c
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+#define av1_apply_selfguided_restoration av1_apply_selfguided_restoration_c
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+#define av1_apply_temporal_filter av1_apply_temporal_filter_c
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define av1_block_error av1_block_error_c
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+#define av1_block_error_lp av1_block_error_lp_c
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+#define av1_build_compound_diffwtd_mask av1_build_compound_diffwtd_mask_c
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+#define av1_build_compound_diffwtd_mask_d16 av1_build_compound_diffwtd_mask_d16_c
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+#define av1_build_compound_diffwtd_mask_highbd av1_build_compound_diffwtd_mask_highbd_c
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+#define av1_calc_indices_dim1 av1_calc_indices_dim1_c
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+#define av1_calc_indices_dim2 av1_calc_indices_dim2_c
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+#define av1_calc_proj_params av1_calc_proj_params_c
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+#define av1_calc_proj_params_high_bd av1_calc_proj_params_high_bd_c
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+#define av1_cnn_convolve_no_maxpool_padding_valid av1_cnn_convolve_no_maxpool_padding_valid_c
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+#define av1_compute_stats av1_compute_stats_c
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+#define av1_compute_stats_highbd av1_compute_stats_highbd_c
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_scale av1_convolve_2d_scale_c
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr av1_convolve_2d_sr_c
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr_intrabc av1_convolve_2d_sr_intrabc_c
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+#define av1_convolve_horiz_rs av1_convolve_horiz_rs_c
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr av1_convolve_x_sr_c
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr_intrabc av1_convolve_x_sr_intrabc_c
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr av1_convolve_y_sr_c
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_dist_wtd_convolve_2d av1_dist_wtd_convolve_2d_c
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+#define av1_dist_wtd_convolve_2d_copy av1_dist_wtd_convolve_2d_copy_c
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_dist_wtd_convolve_x av1_dist_wtd_convolve_x_c
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_dist_wtd_convolve_y av1_dist_wtd_convolve_y_c
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+#define av1_dr_prediction_z1 av1_dr_prediction_z1_c
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+#define av1_dr_prediction_z2 av1_dr_prediction_z2_c
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+#define av1_dr_prediction_z3 av1_dr_prediction_z3_c
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+#define av1_estimate_noise_from_single_plane av1_estimate_noise_from_single_plane_c
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+#define av1_filter_intra_edge av1_filter_intra_edge_c
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+#define av1_filter_intra_predictor av1_filter_intra_predictor_c
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_16x16 av1_fwd_txfm2d_16x16_c
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_16x32 av1_fwd_txfm2d_16x32_c
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_16x4 av1_fwd_txfm2d_16x4_c
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_16x64 av1_fwd_txfm2d_16x64_c
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_16x8 av1_fwd_txfm2d_16x8_c
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_32x16 av1_fwd_txfm2d_32x16_c
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_32x32 av1_fwd_txfm2d_32x32_c
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_32x64 av1_fwd_txfm2d_32x64_c
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_32x8 av1_fwd_txfm2d_32x8_c
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_4x16 av1_fwd_txfm2d_4x16_c
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_4x4 av1_fwd_txfm2d_4x4_c
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_4x8 av1_fwd_txfm2d_4x8_c
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_64x16 av1_fwd_txfm2d_64x16_c
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_64x32 av1_fwd_txfm2d_64x32_c
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_64x64 av1_fwd_txfm2d_64x64_c
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_8x16 av1_fwd_txfm2d_8x16_c
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_8x32 av1_fwd_txfm2d_8x32_c
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_8x4 av1_fwd_txfm2d_8x4_c
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_fwd_txfm2d_8x8 av1_fwd_txfm2d_8x8_c
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define av1_fwht4x4 av1_fwht4x4_c
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+#define av1_get_crc32c_value av1_get_crc32c_value_c
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+#define av1_get_horver_correlation_full av1_get_horver_correlation_full_c
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+#define av1_get_nz_map_contexts av1_get_nz_map_contexts_c
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+#define av1_highbd_apply_temporal_filter av1_highbd_apply_temporal_filter_c
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+#define av1_highbd_block_error av1_highbd_block_error_c
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_c
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_c
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_scale av1_highbd_convolve_2d_scale_c
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr av1_highbd_convolve_2d_sr_c
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr_intrabc av1_highbd_convolve_2d_sr_intrabc_c
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+#define av1_highbd_convolve_horiz_rs av1_highbd_convolve_horiz_rs_c
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr av1_highbd_convolve_x_sr_c
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr_intrabc av1_highbd_convolve_x_sr_intrabc_c
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr av1_highbd_convolve_y_sr_c
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr_intrabc av1_highbd_convolve_y_sr_intrabc_c
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_dist_wtd_convolve_2d av1_highbd_dist_wtd_convolve_2d_c
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+#define av1_highbd_dist_wtd_convolve_2d_copy av1_highbd_dist_wtd_convolve_2d_copy_c
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_dist_wtd_convolve_x av1_highbd_dist_wtd_convolve_x_c
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_dist_wtd_convolve_y av1_highbd_dist_wtd_convolve_y_c
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+#define av1_highbd_dr_prediction_z1 av1_highbd_dr_prediction_z1_c
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+#define av1_highbd_dr_prediction_z2 av1_highbd_dr_prediction_z2_c
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+#define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+#define av1_highbd_estimate_noise_from_single_plane av1_highbd_estimate_noise_from_single_plane_c
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+#define av1_highbd_filter_intra_edge av1_highbd_filter_intra_edge_c
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add av1_highbd_inv_txfm_add_c
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x4 av1_highbd_inv_txfm_add_16x4_c
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_4x16 av1_highbd_inv_txfm_add_4x16_c
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_4x4 av1_highbd_inv_txfm_add_4x4_c
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_4x8 av1_highbd_inv_txfm_add_4x8_c
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x4 av1_highbd_inv_txfm_add_8x4_c
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x8 av1_highbd_inv_txfm_add_8x8_c
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+#define av1_highbd_pixel_proj_error av1_highbd_pixel_proj_error_c
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+#define av1_highbd_quantize_fp av1_highbd_quantize_fp_c
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+#define av1_highbd_upsample_intra_edge av1_highbd_upsample_intra_edge_c
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+#define av1_highbd_warp_affine av1_highbd_warp_affine_c
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+#define av1_highbd_wiener_convolve_add_src av1_highbd_wiener_convolve_add_src_c
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x4 av1_inv_txfm2d_add_4x4_c
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x8 av1_inv_txfm2d_add_8x8_c
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_inv_txfm_add av1_inv_txfm_add_c
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+#define av1_lowbd_fwd_txfm av1_lowbd_fwd_txfm_c
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+#define av1_lowbd_pixel_proj_error av1_lowbd_pixel_proj_error_c
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+#define av1_nn_fast_softmax_16 av1_nn_fast_softmax_16_c
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+#define av1_nn_predict av1_nn_predict_c
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp av1_quantize_fp_c
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp_64x64 av1_quantize_fp_64x64_c
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_lp av1_quantize_lp_c
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+#define av1_resize_and_extend_frame av1_resize_and_extend_frame_c
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+#define av1_round_shift_array av1_round_shift_array_c
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+#define av1_selfguided_restoration av1_selfguided_restoration_c
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+#define av1_txb_init_levels av1_txb_init_levels_c
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+#define av1_upsample_intra_edge av1_upsample_intra_edge_c
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+#define av1_warp_affine av1_warp_affine_c
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+#define av1_wedge_compute_delta_squares av1_wedge_compute_delta_squares_c
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+#define av1_wedge_sign_from_residuals av1_wedge_sign_from_residuals_c
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+#define av1_wedge_sse_from_residuals av1_wedge_sse_from_residuals_c
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+#define av1_wiener_convolve_add_src av1_wiener_convolve_add_src_c
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+#define cdef_copy_rect8_16bit_to_16bit cdef_copy_rect8_16bit_to_16bit_c
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+#define cdef_copy_rect8_8bit_to_16bit cdef_copy_rect8_8bit_to_16bit_c
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_16_0 cdef_filter_16_0_c
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_16_1 cdef_filter_16_1_c
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_16_2 cdef_filter_16_2_c
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_16_3 cdef_filter_16_3_c
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_8_0 cdef_filter_8_0_c
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_8_1 cdef_filter_8_1_c
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_8_2 cdef_filter_8_2_c
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+#define cdef_filter_8_3 cdef_filter_8_3_c
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+#define cdef_find_dir cdef_find_dir_c
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+#define cdef_find_dir_dual cdef_find_dir_dual_c
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+#define cfl_get_luma_subsampling_420_hbd cfl_get_luma_subsampling_420_hbd_c
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+#define cfl_get_luma_subsampling_420_lbd cfl_get_luma_subsampling_420_lbd_c
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+#define cfl_get_luma_subsampling_422_hbd cfl_get_luma_subsampling_422_hbd_c
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+#define cfl_get_luma_subsampling_422_lbd cfl_get_luma_subsampling_422_lbd_c
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+#define cfl_get_luma_subsampling_444_hbd cfl_get_luma_subsampling_444_hbd_c
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+#define cfl_get_luma_subsampling_444_lbd cfl_get_luma_subsampling_444_lbd_c
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+#define cfl_get_predict_hbd_fn cfl_get_predict_hbd_fn_c
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+#define cfl_get_predict_lbd_fn cfl_get_predict_lbd_fn_c
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+#define cfl_get_subtract_average_fn cfl_get_subtract_average_fn_c
+
+void av1_rtcd(void);
+
+#include "config/aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/arm/config/aom_config.asm b/media/libaom/config/linux/arm/config/aom_config.asm
new file mode 100644
index 0000000000..63034fd7e2
--- /dev/null
+++ b/media/libaom/config/linux/arm/config/aom_config.asm
@@ -0,0 +1,93 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+.equ AOM_ARCH_AARCH64, 0
+.equ AOM_ARCH_ARM, 1
+.equ AOM_ARCH_PPC, 0
+.equ AOM_ARCH_X86, 0
+.equ AOM_ARCH_X86_64, 0
+.equ CONFIG_ACCOUNTING, 0
+.equ CONFIG_ANALYZER, 0
+.equ CONFIG_AV1_DECODER, 1
+.equ CONFIG_AV1_ENCODER, 1
+.equ CONFIG_AV1_HIGHBITDEPTH, 1
+.equ CONFIG_AV1_TEMPORAL_DENOISING, 0
+.equ CONFIG_BIG_ENDIAN, 0
+.equ CONFIG_BITRATE_ACCURACY, 0
+.equ CONFIG_BITRATE_ACCURACY_BL, 0
+.equ CONFIG_BITSTREAM_DEBUG, 0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING, 0
+.equ CONFIG_COLLECT_COMPONENT_TIMING, 0
+.equ CONFIG_COLLECT_PARTITION_STATS, 0
+.equ CONFIG_COLLECT_RD_STATS, 0
+.equ CONFIG_CWG_C013, 0
+.equ CONFIG_DEBUG, 0
+.equ CONFIG_DENOISE, 1
+.equ CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8, 1
+.equ CONFIG_ENTROPY_STATS, 0
+.equ CONFIG_EXCLUDE_SIMD_MISMATCH, 0
+.equ CONFIG_FPMT_TEST, 0
+.equ CONFIG_GCC, 1
+.equ CONFIG_GCOV, 0
+.equ CONFIG_GPROF, 0
+.equ CONFIG_INSPECTION, 0
+.equ CONFIG_INTERNAL_STATS, 0
+.equ CONFIG_INTER_STATS_ONLY, 0
+.equ CONFIG_LIBVMAF_PSNR_PEAK, 1
+.equ CONFIG_LIBYUV, 0
+.equ CONFIG_MAX_DECODE_PROFILE, 2
+.equ CONFIG_MISMATCH_DEBUG, 0
+.equ CONFIG_MULTITHREAD, 1
+.equ CONFIG_NN_V2, 0
+.equ CONFIG_NORMAL_TILE_MODE, 0
+.equ CONFIG_OPTICAL_FLOW_API, 0
+.equ CONFIG_OS_SUPPORT, 1
+.equ CONFIG_OUTPUT_FRAME_SIZE, 0
+.equ CONFIG_PARTITION_SEARCH_ORDER, 0
+.equ CONFIG_PIC, 1
+.equ CONFIG_RATECTRL_LOG, 0
+.equ CONFIG_RD_COMMAND, 0
+.equ CONFIG_RD_DEBUG, 0
+.equ CONFIG_REALTIME_ONLY, 0
+.equ CONFIG_RT_ML_PARTITIONING, 0
+.equ CONFIG_RUNTIME_CPU_DETECT, 1
+.equ CONFIG_SALIENCY_MAP, 0
+.equ CONFIG_SHARED, 0
+.equ CONFIG_SIZE_LIMIT, 0
+.equ CONFIG_SPATIAL_RESAMPLING, 1
+.equ CONFIG_SPEED_STATS, 0
+.equ CONFIG_TFLITE, 0
+.equ CONFIG_THREE_PASS, 0
+.equ CONFIG_TUNE_BUTTERAUGLI, 0
+.equ CONFIG_TUNE_VMAF, 0
+.equ CONFIG_WEBM_IO, 0
+.equ DECODE_HEIGHT_LIMIT, 0
+.equ DECODE_WIDTH_LIMIT, 0
+.equ FORCE_HIGHBITDEPTH_DECODING, 0
+.equ HAVE_ARM_CRC32, 0
+.equ HAVE_AVX, 0
+.equ HAVE_AVX2, 0
+.equ HAVE_FEXCEPT, 1
+.equ HAVE_MMX, 0
+.equ HAVE_NEON, 1
+.equ HAVE_NEON_DOTPROD, 0
+.equ HAVE_NEON_I8MM, 0
+.equ HAVE_SSE, 0
+.equ HAVE_SSE2, 0
+.equ HAVE_SSE3, 0
+.equ HAVE_SSE4_1, 0
+.equ HAVE_SSE4_2, 0
+.equ HAVE_SSSE3, 0
+.equ HAVE_SVE, 0
+.equ HAVE_VSX, 0
+.equ HAVE_WXWIDGETS, 0
+.equ STATIC_LINK_JXL, 0
+.section	.note.GNU-stack,"",%progbits
\ No newline at end of file
diff --git a/media/libaom/config/linux/arm/config/aom_config.h b/media/libaom/config/linux/arm/config/aom_config.h
new file mode 100644
index 0000000000..3cbe7bf169
--- /dev/null
+++ b/media/libaom/config/linux/arm/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 1
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 0
+#define AOM_ARCH_X86_64 0
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 1
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 0
+#define HAVE_NEON 1
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_SSE4_2 0
+#define HAVE_SSSE3 0
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..50ee78932c
--- /dev/null
+++ b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
@@ -0,0 +1,9275 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_neon(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_4x4)(const uint8_t *, int p);
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_neon(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_8x8)(const uint8_t *, int p);
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_neon(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_hmask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_vmask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_pred_neon(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_neon(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_neon(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_dist_wtd_comp_avg_pred_neon(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x128_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x128_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_neon(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+void aom_fdct4x4_lp_neon(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4_lp)(const int16_t *input, int16_t *output, int stride);
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+#define aom_fft16x16_float aom_fft16x16_float_c
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+#define aom_fft32x32_float aom_fft32x32_float_c
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+#define aom_fft4x4_float aom_fft4x4_float_c
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+#define aom_fft8x8_float aom_fft8x8_float_c
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_neon(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+RTCD_EXTERN void (*aom_get_blk_sse_sum)(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_neon(const int16_t *);
+RTCD_EXTERN unsigned int (*aom_get_mb_ss)(const int16_t *);
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+RTCD_EXTERN void (*aom_get_var_sse_sum_16x16_dual)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+RTCD_EXTERN void (*aom_get_var_sse_sum_8x8_quad)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_4x4_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_4x4)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_neon(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8_dual)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+RTCD_EXTERN uint32_t (*aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_highbd_avg_4x4_neon(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_highbd_avg_4x4)(const uint8_t *, int p);
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_highbd_avg_8x8_neon(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_highbd_avg_8x8)(const uint8_t *, int p);
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_neon(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_hmask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+void aom_highbd_blend_a64_mask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_vmask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+void aom_highbd_comp_avg_pred_neon(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+RTCD_EXTERN void (*aom_highbd_comp_avg_pred)(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_neon(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_highbd_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_neon(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_highbd_dist_wtd_comp_avg_pred_neon(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_highbd_dist_wtd_comp_avg_pred)(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_16x16_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_32x32_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_14_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_6_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_6_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_14_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_6_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_6_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_neon(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_highbd_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+RTCD_EXTERN void (*aom_highbd_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_paeth_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_adaptive_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x128_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x128_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_h_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_smooth_v_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_neon(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_highbd_sse)(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_highbd_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_highbd_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+#define aom_ifft16x16_float aom_ifft16x16_float_c
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+#define aom_ifft32x32_float aom_ifft32x32_float_c
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+#define aom_ifft4x4_float aom_ifft4x4_float_c
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+#define aom_ifft8x8_float aom_ifft8x8_float_c
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_neon(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_col)(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_neon(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_neon(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_14_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_14_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_4_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_6_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_6_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_14_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_14_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_4_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_6_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_6_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_6_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_8_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x128x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x64x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x16x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x32x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x4x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x64x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x8x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x16x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x32x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x64x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x8x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x16x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x4x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x8x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x128x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x16x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x32x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x64x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x16x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x32x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x4x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_neon(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x8x4d_neon(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x128_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x32_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+RTCD_EXTERN void (*aom_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_neon(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_16xh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_neon(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_neon(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit_highbd)(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x128_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x64_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x16_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x32_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x4_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x8_neon(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_neon(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32_adaptive aom_quantize_b_32x32_adaptive_c
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64_adaptive aom_quantize_b_64x64_adaptive_c
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_adaptive aom_quantize_b_adaptive_c
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x32_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x3d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x4x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x8x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+int aom_satd_neon(const tran_low_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const tran_low_t *coeff, int length);
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+int aom_satd_lp_neon(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd_lp)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+void aom_scaled_2d_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_neon(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_sse)(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_c
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_neon(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_neon(const int16_t *src, uint32_t N);
+RTCD_EXTERN uint64_t (*aom_sum_squares_i16)(const int16_t *src, uint32_t N);
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_neon(const int16_t *src, int src_stride, int width, int height, int *sum);
+RTCD_EXTERN uint64_t (*aom_sum_sse_2d_i16)(const int16_t *src, int src_stride, int width, int height, int *sum);
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_neon(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u16)(uint8_t *src, int src_stride, int width, int height);
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_neon(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u8)(uint8_t *src, int src_stride, int width, int height);
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_neon(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+#define av1_compute_cross_correlation av1_compute_cross_correlation_c
+
+void aom_dsp_rtcd(void);
+
+#include "config/aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = aom_arm_cpu_caps();
+
+    (void)flags;
+
+    aom_avg_4x4 = aom_avg_4x4_c;
+    if (flags & HAS_NEON) aom_avg_4x4 = aom_avg_4x4_neon;
+    aom_avg_8x8 = aom_avg_8x8_c;
+    if (flags & HAS_NEON) aom_avg_8x8 = aom_avg_8x8_neon;
+    aom_avg_8x8_quad = aom_avg_8x8_quad_c;
+    if (flags & HAS_NEON) aom_avg_8x8_quad = aom_avg_8x8_quad_neon;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_NEON) aom_blend_a64_hmask = aom_blend_a64_hmask_neon;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_NEON) aom_blend_a64_mask = aom_blend_a64_mask_neon;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_NEON) aom_blend_a64_vmask = aom_blend_a64_vmask_neon;
+    aom_comp_avg_pred = aom_comp_avg_pred_c;
+    if (flags & HAS_NEON) aom_comp_avg_pred = aom_comp_avg_pred_neon;
+    aom_comp_mask_pred = aom_comp_mask_pred_c;
+    if (flags & HAS_NEON) aom_comp_mask_pred = aom_comp_mask_pred_neon;
+    aom_compute_flow_at_point = aom_compute_flow_at_point_c;
+    if (flags & HAS_NEON) aom_compute_flow_at_point = aom_compute_flow_at_point_neon;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
+    if (flags & HAS_NEON) aom_convolve8_horiz = aom_convolve8_horiz_neon;
+    aom_convolve8_vert = aom_convolve8_vert_c;
+    if (flags & HAS_NEON) aom_convolve8_vert = aom_convolve8_vert_neon;
+    aom_convolve_copy = aom_convolve_copy_c;
+    if (flags & HAS_NEON) aom_convolve_copy = aom_convolve_copy_neon;
+    aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_neon;
+    aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_neon;
+    aom_dc_128_predictor_16x4 = aom_dc_128_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_16x4 = aom_dc_128_predictor_16x4_neon;
+    aom_dc_128_predictor_16x64 = aom_dc_128_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_16x64 = aom_dc_128_predictor_16x64_neon;
+    aom_dc_128_predictor_16x8 = aom_dc_128_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_16x8 = aom_dc_128_predictor_16x8_neon;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_neon;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_neon;
+    aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_neon;
+    aom_dc_128_predictor_32x8 = aom_dc_128_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_32x8 = aom_dc_128_predictor_32x8_neon;
+    aom_dc_128_predictor_4x16 = aom_dc_128_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_4x16 = aom_dc_128_predictor_4x16_neon;
+    aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_neon;
+    aom_dc_128_predictor_4x8 = aom_dc_128_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_4x8 = aom_dc_128_predictor_4x8_neon;
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_neon;
+    aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_neon;
+    aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_neon;
+    aom_dc_128_predictor_8x16 = aom_dc_128_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_8x16 = aom_dc_128_predictor_8x16_neon;
+    aom_dc_128_predictor_8x32 = aom_dc_128_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_8x32 = aom_dc_128_predictor_8x32_neon;
+    aom_dc_128_predictor_8x4 = aom_dc_128_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_8x4 = aom_dc_128_predictor_8x4_neon;
+    aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_neon;
+    aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_neon;
+    aom_dc_left_predictor_16x32 = aom_dc_left_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_16x32 = aom_dc_left_predictor_16x32_neon;
+    aom_dc_left_predictor_16x4 = aom_dc_left_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_16x4 = aom_dc_left_predictor_16x4_neon;
+    aom_dc_left_predictor_16x64 = aom_dc_left_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_16x64 = aom_dc_left_predictor_16x64_neon;
+    aom_dc_left_predictor_16x8 = aom_dc_left_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_16x8 = aom_dc_left_predictor_16x8_neon;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_neon;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_neon;
+    aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_neon;
+    aom_dc_left_predictor_32x8 = aom_dc_left_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_32x8 = aom_dc_left_predictor_32x8_neon;
+    aom_dc_left_predictor_4x16 = aom_dc_left_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_4x16 = aom_dc_left_predictor_4x16_neon;
+    aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_neon;
+    aom_dc_left_predictor_4x8 = aom_dc_left_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_4x8 = aom_dc_left_predictor_4x8_neon;
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_neon;
+    aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_neon;
+    aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_neon;
+    aom_dc_left_predictor_8x16 = aom_dc_left_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_8x16 = aom_dc_left_predictor_8x16_neon;
+    aom_dc_left_predictor_8x32 = aom_dc_left_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_8x32 = aom_dc_left_predictor_8x32_neon;
+    aom_dc_left_predictor_8x4 = aom_dc_left_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_8x4 = aom_dc_left_predictor_8x4_neon;
+    aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_neon;
+    aom_dc_predictor_16x16 = aom_dc_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_predictor_16x16 = aom_dc_predictor_16x16_neon;
+    aom_dc_predictor_16x32 = aom_dc_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_dc_predictor_16x32 = aom_dc_predictor_16x32_neon;
+    aom_dc_predictor_16x4 = aom_dc_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_dc_predictor_16x4 = aom_dc_predictor_16x4_neon;
+    aom_dc_predictor_16x64 = aom_dc_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_dc_predictor_16x64 = aom_dc_predictor_16x64_neon;
+    aom_dc_predictor_16x8 = aom_dc_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_dc_predictor_16x8 = aom_dc_predictor_16x8_neon;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_neon;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_neon;
+    aom_dc_predictor_32x64 = aom_dc_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_neon;
+    aom_dc_predictor_32x8 = aom_dc_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_dc_predictor_32x8 = aom_dc_predictor_32x8_neon;
+    aom_dc_predictor_4x16 = aom_dc_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_dc_predictor_4x16 = aom_dc_predictor_4x16_neon;
+    aom_dc_predictor_4x4 = aom_dc_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_predictor_4x4 = aom_dc_predictor_4x4_neon;
+    aom_dc_predictor_4x8 = aom_dc_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_dc_predictor_4x8 = aom_dc_predictor_4x8_neon;
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_neon;
+    aom_dc_predictor_64x32 = aom_dc_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_neon;
+    aom_dc_predictor_64x64 = aom_dc_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_neon;
+    aom_dc_predictor_8x16 = aom_dc_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_dc_predictor_8x16 = aom_dc_predictor_8x16_neon;
+    aom_dc_predictor_8x32 = aom_dc_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_dc_predictor_8x32 = aom_dc_predictor_8x32_neon;
+    aom_dc_predictor_8x4 = aom_dc_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_dc_predictor_8x4 = aom_dc_predictor_8x4_neon;
+    aom_dc_predictor_8x8 = aom_dc_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_predictor_8x8 = aom_dc_predictor_8x8_neon;
+    aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_neon;
+    aom_dc_top_predictor_16x32 = aom_dc_top_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_16x32 = aom_dc_top_predictor_16x32_neon;
+    aom_dc_top_predictor_16x4 = aom_dc_top_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_16x4 = aom_dc_top_predictor_16x4_neon;
+    aom_dc_top_predictor_16x64 = aom_dc_top_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_16x64 = aom_dc_top_predictor_16x64_neon;
+    aom_dc_top_predictor_16x8 = aom_dc_top_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_16x8 = aom_dc_top_predictor_16x8_neon;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_neon;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_neon;
+    aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_neon;
+    aom_dc_top_predictor_32x8 = aom_dc_top_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_32x8 = aom_dc_top_predictor_32x8_neon;
+    aom_dc_top_predictor_4x16 = aom_dc_top_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_4x16 = aom_dc_top_predictor_4x16_neon;
+    aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_neon;
+    aom_dc_top_predictor_4x8 = aom_dc_top_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_4x8 = aom_dc_top_predictor_4x8_neon;
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_neon;
+    aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_neon;
+    aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_neon;
+    aom_dc_top_predictor_8x16 = aom_dc_top_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_8x16 = aom_dc_top_predictor_8x16_neon;
+    aom_dc_top_predictor_8x32 = aom_dc_top_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_8x32 = aom_dc_top_predictor_8x32_neon;
+    aom_dc_top_predictor_8x4 = aom_dc_top_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_8x4 = aom_dc_top_predictor_8x4_neon;
+    aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_neon;
+    aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_NEON) aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_neon;
+    aom_dist_wtd_sad128x128_avg = aom_dist_wtd_sad128x128_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad128x128_avg = aom_dist_wtd_sad128x128_avg_neon;
+    aom_dist_wtd_sad128x64_avg = aom_dist_wtd_sad128x64_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad128x64_avg = aom_dist_wtd_sad128x64_avg_neon;
+    aom_dist_wtd_sad16x16_avg = aom_dist_wtd_sad16x16_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad16x16_avg = aom_dist_wtd_sad16x16_avg_neon;
+    aom_dist_wtd_sad16x32_avg = aom_dist_wtd_sad16x32_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad16x32_avg = aom_dist_wtd_sad16x32_avg_neon;
+    aom_dist_wtd_sad16x4_avg = aom_dist_wtd_sad16x4_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad16x4_avg = aom_dist_wtd_sad16x4_avg_neon;
+    aom_dist_wtd_sad16x64_avg = aom_dist_wtd_sad16x64_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad16x64_avg = aom_dist_wtd_sad16x64_avg_neon;
+    aom_dist_wtd_sad16x8_avg = aom_dist_wtd_sad16x8_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad16x8_avg = aom_dist_wtd_sad16x8_avg_neon;
+    aom_dist_wtd_sad32x16_avg = aom_dist_wtd_sad32x16_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad32x16_avg = aom_dist_wtd_sad32x16_avg_neon;
+    aom_dist_wtd_sad32x32_avg = aom_dist_wtd_sad32x32_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad32x32_avg = aom_dist_wtd_sad32x32_avg_neon;
+    aom_dist_wtd_sad32x64_avg = aom_dist_wtd_sad32x64_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad32x64_avg = aom_dist_wtd_sad32x64_avg_neon;
+    aom_dist_wtd_sad32x8_avg = aom_dist_wtd_sad32x8_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad32x8_avg = aom_dist_wtd_sad32x8_avg_neon;
+    aom_dist_wtd_sad4x16_avg = aom_dist_wtd_sad4x16_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad4x16_avg = aom_dist_wtd_sad4x16_avg_neon;
+    aom_dist_wtd_sad4x4_avg = aom_dist_wtd_sad4x4_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad4x4_avg = aom_dist_wtd_sad4x4_avg_neon;
+    aom_dist_wtd_sad4x8_avg = aom_dist_wtd_sad4x8_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad4x8_avg = aom_dist_wtd_sad4x8_avg_neon;
+    aom_dist_wtd_sad64x128_avg = aom_dist_wtd_sad64x128_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad64x128_avg = aom_dist_wtd_sad64x128_avg_neon;
+    aom_dist_wtd_sad64x16_avg = aom_dist_wtd_sad64x16_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad64x16_avg = aom_dist_wtd_sad64x16_avg_neon;
+    aom_dist_wtd_sad64x32_avg = aom_dist_wtd_sad64x32_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad64x32_avg = aom_dist_wtd_sad64x32_avg_neon;
+    aom_dist_wtd_sad64x64_avg = aom_dist_wtd_sad64x64_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad64x64_avg = aom_dist_wtd_sad64x64_avg_neon;
+    aom_dist_wtd_sad8x16_avg = aom_dist_wtd_sad8x16_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad8x16_avg = aom_dist_wtd_sad8x16_avg_neon;
+    aom_dist_wtd_sad8x32_avg = aom_dist_wtd_sad8x32_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad8x32_avg = aom_dist_wtd_sad8x32_avg_neon;
+    aom_dist_wtd_sad8x4_avg = aom_dist_wtd_sad8x4_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad8x4_avg = aom_dist_wtd_sad8x4_avg_neon;
+    aom_dist_wtd_sad8x8_avg = aom_dist_wtd_sad8x8_avg_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sad8x8_avg = aom_dist_wtd_sad8x8_avg_neon;
+    aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_neon;
+    aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_neon;
+    aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_neon;
+    aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_neon;
+    aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_neon;
+    aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_neon;
+    aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_neon;
+    aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_neon;
+    aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_neon;
+    aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_neon;
+    aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_neon;
+    aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_neon;
+    aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_neon;
+    aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_neon;
+    aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_neon;
+    aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_neon;
+    aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_neon;
+    aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_neon;
+    aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_neon;
+    aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_neon;
+    aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_neon;
+    aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_neon;
+    aom_fdct4x4 = aom_fdct4x4_c;
+    if (flags & HAS_NEON) aom_fdct4x4 = aom_fdct4x4_neon;
+    aom_fdct4x4_lp = aom_fdct4x4_lp_c;
+    if (flags & HAS_NEON) aom_fdct4x4_lp = aom_fdct4x4_lp_neon;
+    aom_get_blk_sse_sum = aom_get_blk_sse_sum_c;
+    if (flags & HAS_NEON) aom_get_blk_sse_sum = aom_get_blk_sse_sum_neon;
+    aom_get_mb_ss = aom_get_mb_ss_c;
+    if (flags & HAS_NEON) aom_get_mb_ss = aom_get_mb_ss_neon;
+    aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_c;
+    if (flags & HAS_NEON) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_neon;
+    aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_c;
+    if (flags & HAS_NEON) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_neon;
+    aom_h_predictor_16x16 = aom_h_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_h_predictor_16x16 = aom_h_predictor_16x16_neon;
+    aom_h_predictor_16x32 = aom_h_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_h_predictor_16x32 = aom_h_predictor_16x32_neon;
+    aom_h_predictor_16x4 = aom_h_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_h_predictor_16x4 = aom_h_predictor_16x4_neon;
+    aom_h_predictor_16x64 = aom_h_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_h_predictor_16x64 = aom_h_predictor_16x64_neon;
+    aom_h_predictor_16x8 = aom_h_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_h_predictor_16x8 = aom_h_predictor_16x8_neon;
+    aom_h_predictor_32x16 = aom_h_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_h_predictor_32x16 = aom_h_predictor_32x16_neon;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_h_predictor_32x32 = aom_h_predictor_32x32_neon;
+    aom_h_predictor_32x64 = aom_h_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_h_predictor_32x64 = aom_h_predictor_32x64_neon;
+    aom_h_predictor_32x8 = aom_h_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_h_predictor_32x8 = aom_h_predictor_32x8_neon;
+    aom_h_predictor_4x16 = aom_h_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_h_predictor_4x16 = aom_h_predictor_4x16_neon;
+    aom_h_predictor_4x4 = aom_h_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_h_predictor_4x4 = aom_h_predictor_4x4_neon;
+    aom_h_predictor_4x8 = aom_h_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_h_predictor_4x8 = aom_h_predictor_4x8_neon;
+    aom_h_predictor_64x16 = aom_h_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_h_predictor_64x16 = aom_h_predictor_64x16_neon;
+    aom_h_predictor_64x32 = aom_h_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_h_predictor_64x32 = aom_h_predictor_64x32_neon;
+    aom_h_predictor_64x64 = aom_h_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_h_predictor_64x64 = aom_h_predictor_64x64_neon;
+    aom_h_predictor_8x16 = aom_h_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_h_predictor_8x16 = aom_h_predictor_8x16_neon;
+    aom_h_predictor_8x32 = aom_h_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_h_predictor_8x32 = aom_h_predictor_8x32_neon;
+    aom_h_predictor_8x4 = aom_h_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_h_predictor_8x4 = aom_h_predictor_8x4_neon;
+    aom_h_predictor_8x8 = aom_h_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_h_predictor_8x8 = aom_h_predictor_8x8_neon;
+    aom_hadamard_16x16 = aom_hadamard_16x16_c;
+    if (flags & HAS_NEON) aom_hadamard_16x16 = aom_hadamard_16x16_neon;
+    aom_hadamard_32x32 = aom_hadamard_32x32_c;
+    if (flags & HAS_NEON) aom_hadamard_32x32 = aom_hadamard_32x32_neon;
+    aom_hadamard_4x4 = aom_hadamard_4x4_c;
+    if (flags & HAS_NEON) aom_hadamard_4x4 = aom_hadamard_4x4_neon;
+    aom_hadamard_8x8 = aom_hadamard_8x8_c;
+    if (flags & HAS_NEON) aom_hadamard_8x8 = aom_hadamard_8x8_neon;
+    aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_c;
+    if (flags & HAS_NEON) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_neon;
+    aom_hadamard_lp_8x8 = aom_hadamard_lp_8x8_c;
+    if (flags & HAS_NEON) aom_hadamard_lp_8x8 = aom_hadamard_lp_8x8_neon;
+    aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_c;
+    if (flags & HAS_NEON) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_neon;
+    aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 = aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_neon;
+    aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_neon;
+    aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_neon;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_neon;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_neon;
+    aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_neon;
+    aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_neon;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_neon;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_neon;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_neon;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_neon;
+    aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_neon;
+    aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_neon;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_neon;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_neon;
+    aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_neon;
+    aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_neon;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_neon;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_neon;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_neon;
+    aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_neon;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_neon;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_neon;
+    aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_neon;
+    aom_highbd_10_mse16x8 = aom_highbd_10_mse16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_mse16x8 = aom_highbd_10_mse16x8_neon;
+    aom_highbd_10_mse8x16 = aom_highbd_10_mse8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_mse8x16 = aom_highbd_10_mse8x16_neon;
+    aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_neon;
+    aom_highbd_10_obmc_sub_pixel_variance128x128 = aom_highbd_10_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance128x128 = aom_highbd_10_obmc_sub_pixel_variance128x128_neon;
+    aom_highbd_10_obmc_sub_pixel_variance128x64 = aom_highbd_10_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance128x64 = aom_highbd_10_obmc_sub_pixel_variance128x64_neon;
+    aom_highbd_10_obmc_sub_pixel_variance16x16 = aom_highbd_10_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance16x16 = aom_highbd_10_obmc_sub_pixel_variance16x16_neon;
+    aom_highbd_10_obmc_sub_pixel_variance16x32 = aom_highbd_10_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance16x32 = aom_highbd_10_obmc_sub_pixel_variance16x32_neon;
+    aom_highbd_10_obmc_sub_pixel_variance16x4 = aom_highbd_10_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance16x4 = aom_highbd_10_obmc_sub_pixel_variance16x4_neon;
+    aom_highbd_10_obmc_sub_pixel_variance16x64 = aom_highbd_10_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance16x64 = aom_highbd_10_obmc_sub_pixel_variance16x64_neon;
+    aom_highbd_10_obmc_sub_pixel_variance16x8 = aom_highbd_10_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance16x8 = aom_highbd_10_obmc_sub_pixel_variance16x8_neon;
+    aom_highbd_10_obmc_sub_pixel_variance32x16 = aom_highbd_10_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance32x16 = aom_highbd_10_obmc_sub_pixel_variance32x16_neon;
+    aom_highbd_10_obmc_sub_pixel_variance32x32 = aom_highbd_10_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance32x32 = aom_highbd_10_obmc_sub_pixel_variance32x32_neon;
+    aom_highbd_10_obmc_sub_pixel_variance32x64 = aom_highbd_10_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance32x64 = aom_highbd_10_obmc_sub_pixel_variance32x64_neon;
+    aom_highbd_10_obmc_sub_pixel_variance32x8 = aom_highbd_10_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance32x8 = aom_highbd_10_obmc_sub_pixel_variance32x8_neon;
+    aom_highbd_10_obmc_sub_pixel_variance4x16 = aom_highbd_10_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance4x16 = aom_highbd_10_obmc_sub_pixel_variance4x16_neon;
+    aom_highbd_10_obmc_sub_pixel_variance4x4 = aom_highbd_10_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance4x4 = aom_highbd_10_obmc_sub_pixel_variance4x4_neon;
+    aom_highbd_10_obmc_sub_pixel_variance4x8 = aom_highbd_10_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance4x8 = aom_highbd_10_obmc_sub_pixel_variance4x8_neon;
+    aom_highbd_10_obmc_sub_pixel_variance64x128 = aom_highbd_10_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance64x128 = aom_highbd_10_obmc_sub_pixel_variance64x128_neon;
+    aom_highbd_10_obmc_sub_pixel_variance64x16 = aom_highbd_10_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance64x16 = aom_highbd_10_obmc_sub_pixel_variance64x16_neon;
+    aom_highbd_10_obmc_sub_pixel_variance64x32 = aom_highbd_10_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance64x32 = aom_highbd_10_obmc_sub_pixel_variance64x32_neon;
+    aom_highbd_10_obmc_sub_pixel_variance64x64 = aom_highbd_10_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance64x64 = aom_highbd_10_obmc_sub_pixel_variance64x64_neon;
+    aom_highbd_10_obmc_sub_pixel_variance8x16 = aom_highbd_10_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance8x16 = aom_highbd_10_obmc_sub_pixel_variance8x16_neon;
+    aom_highbd_10_obmc_sub_pixel_variance8x32 = aom_highbd_10_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance8x32 = aom_highbd_10_obmc_sub_pixel_variance8x32_neon;
+    aom_highbd_10_obmc_sub_pixel_variance8x4 = aom_highbd_10_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance8x4 = aom_highbd_10_obmc_sub_pixel_variance8x4_neon;
+    aom_highbd_10_obmc_sub_pixel_variance8x8 = aom_highbd_10_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_sub_pixel_variance8x8 = aom_highbd_10_obmc_sub_pixel_variance8x8_neon;
+    aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_neon;
+    aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_neon;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_neon;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_neon;
+    aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_neon;
+    aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_neon;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_neon;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_neon;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_neon;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_neon;
+    aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_neon;
+    aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_neon;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_neon;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_neon;
+    aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_neon;
+    aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_neon;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_neon;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_neon;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_neon;
+    aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_neon;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_neon;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_neon;
+    aom_highbd_10_sub_pixel_avg_variance128x128 = aom_highbd_10_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance128x128 = aom_highbd_10_sub_pixel_avg_variance128x128_neon;
+    aom_highbd_10_sub_pixel_avg_variance128x64 = aom_highbd_10_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance128x64 = aom_highbd_10_sub_pixel_avg_variance128x64_neon;
+    aom_highbd_10_sub_pixel_avg_variance16x16 = aom_highbd_10_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance16x16 = aom_highbd_10_sub_pixel_avg_variance16x16_neon;
+    aom_highbd_10_sub_pixel_avg_variance16x32 = aom_highbd_10_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance16x32 = aom_highbd_10_sub_pixel_avg_variance16x32_neon;
+    aom_highbd_10_sub_pixel_avg_variance16x4 = aom_highbd_10_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance16x4 = aom_highbd_10_sub_pixel_avg_variance16x4_neon;
+    aom_highbd_10_sub_pixel_avg_variance16x64 = aom_highbd_10_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance16x64 = aom_highbd_10_sub_pixel_avg_variance16x64_neon;
+    aom_highbd_10_sub_pixel_avg_variance16x8 = aom_highbd_10_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance16x8 = aom_highbd_10_sub_pixel_avg_variance16x8_neon;
+    aom_highbd_10_sub_pixel_avg_variance32x16 = aom_highbd_10_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance32x16 = aom_highbd_10_sub_pixel_avg_variance32x16_neon;
+    aom_highbd_10_sub_pixel_avg_variance32x32 = aom_highbd_10_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance32x32 = aom_highbd_10_sub_pixel_avg_variance32x32_neon;
+    aom_highbd_10_sub_pixel_avg_variance32x64 = aom_highbd_10_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance32x64 = aom_highbd_10_sub_pixel_avg_variance32x64_neon;
+    aom_highbd_10_sub_pixel_avg_variance32x8 = aom_highbd_10_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance32x8 = aom_highbd_10_sub_pixel_avg_variance32x8_neon;
+    aom_highbd_10_sub_pixel_avg_variance4x16 = aom_highbd_10_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance4x16 = aom_highbd_10_sub_pixel_avg_variance4x16_neon;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_neon;
+    aom_highbd_10_sub_pixel_avg_variance4x8 = aom_highbd_10_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance4x8 = aom_highbd_10_sub_pixel_avg_variance4x8_neon;
+    aom_highbd_10_sub_pixel_avg_variance64x128 = aom_highbd_10_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance64x128 = aom_highbd_10_sub_pixel_avg_variance64x128_neon;
+    aom_highbd_10_sub_pixel_avg_variance64x16 = aom_highbd_10_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance64x16 = aom_highbd_10_sub_pixel_avg_variance64x16_neon;
+    aom_highbd_10_sub_pixel_avg_variance64x32 = aom_highbd_10_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance64x32 = aom_highbd_10_sub_pixel_avg_variance64x32_neon;
+    aom_highbd_10_sub_pixel_avg_variance64x64 = aom_highbd_10_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance64x64 = aom_highbd_10_sub_pixel_avg_variance64x64_neon;
+    aom_highbd_10_sub_pixel_avg_variance8x16 = aom_highbd_10_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance8x16 = aom_highbd_10_sub_pixel_avg_variance8x16_neon;
+    aom_highbd_10_sub_pixel_avg_variance8x32 = aom_highbd_10_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance8x32 = aom_highbd_10_sub_pixel_avg_variance8x32_neon;
+    aom_highbd_10_sub_pixel_avg_variance8x4 = aom_highbd_10_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance8x4 = aom_highbd_10_sub_pixel_avg_variance8x4_neon;
+    aom_highbd_10_sub_pixel_avg_variance8x8 = aom_highbd_10_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_avg_variance8x8 = aom_highbd_10_sub_pixel_avg_variance8x8_neon;
+    aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_neon;
+    aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_neon;
+    aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_neon;
+    aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_neon;
+    aom_highbd_10_sub_pixel_variance16x4 = aom_highbd_10_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance16x4 = aom_highbd_10_sub_pixel_variance16x4_neon;
+    aom_highbd_10_sub_pixel_variance16x64 = aom_highbd_10_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance16x64 = aom_highbd_10_sub_pixel_variance16x64_neon;
+    aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_neon;
+    aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_neon;
+    aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_neon;
+    aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_neon;
+    aom_highbd_10_sub_pixel_variance32x8 = aom_highbd_10_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance32x8 = aom_highbd_10_sub_pixel_variance32x8_neon;
+    aom_highbd_10_sub_pixel_variance4x16 = aom_highbd_10_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance4x16 = aom_highbd_10_sub_pixel_variance4x16_neon;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_neon;
+    aom_highbd_10_sub_pixel_variance4x8 = aom_highbd_10_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance4x8 = aom_highbd_10_sub_pixel_variance4x8_neon;
+    aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_neon;
+    aom_highbd_10_sub_pixel_variance64x16 = aom_highbd_10_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance64x16 = aom_highbd_10_sub_pixel_variance64x16_neon;
+    aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_neon;
+    aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_neon;
+    aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_neon;
+    aom_highbd_10_sub_pixel_variance8x32 = aom_highbd_10_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance8x32 = aom_highbd_10_sub_pixel_variance8x32_neon;
+    aom_highbd_10_sub_pixel_variance8x4 = aom_highbd_10_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance8x4 = aom_highbd_10_sub_pixel_variance8x4_neon;
+    aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_neon;
+    aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_neon;
+    aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_neon;
+    aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_neon;
+    aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_neon;
+    aom_highbd_10_variance16x4 = aom_highbd_10_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance16x4 = aom_highbd_10_variance16x4_neon;
+    aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_neon;
+    aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_neon;
+    aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_neon;
+    aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_neon;
+    aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_neon;
+    aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_neon;
+    aom_highbd_10_variance4x16 = aom_highbd_10_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance4x16 = aom_highbd_10_variance4x16_neon;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_neon;
+    aom_highbd_10_variance4x8 = aom_highbd_10_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance4x8 = aom_highbd_10_variance4x8_neon;
+    aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_neon;
+    aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_neon;
+    aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_neon;
+    aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_neon;
+    aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_neon;
+    aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_neon;
+    aom_highbd_10_variance8x4 = aom_highbd_10_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance8x4 = aom_highbd_10_variance8x4_neon;
+    aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_neon;
+    aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 = aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_neon;
+    aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_neon;
+    aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_neon;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_neon;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_neon;
+    aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_neon;
+    aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_neon;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_neon;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_neon;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_neon;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_neon;
+    aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_neon;
+    aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_neon;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_neon;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_neon;
+    aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_neon;
+    aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_neon;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_neon;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_neon;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_neon;
+    aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_neon;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_neon;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_neon;
+    aom_highbd_12_mse16x16 = aom_highbd_12_mse16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_mse16x16 = aom_highbd_12_mse16x16_neon;
+    aom_highbd_12_mse16x8 = aom_highbd_12_mse16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_mse16x8 = aom_highbd_12_mse16x8_neon;
+    aom_highbd_12_mse8x16 = aom_highbd_12_mse8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_mse8x16 = aom_highbd_12_mse8x16_neon;
+    aom_highbd_12_mse8x8 = aom_highbd_12_mse8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_mse8x8 = aom_highbd_12_mse8x8_neon;
+    aom_highbd_12_obmc_sub_pixel_variance128x128 = aom_highbd_12_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance128x128 = aom_highbd_12_obmc_sub_pixel_variance128x128_neon;
+    aom_highbd_12_obmc_sub_pixel_variance128x64 = aom_highbd_12_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance128x64 = aom_highbd_12_obmc_sub_pixel_variance128x64_neon;
+    aom_highbd_12_obmc_sub_pixel_variance16x16 = aom_highbd_12_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance16x16 = aom_highbd_12_obmc_sub_pixel_variance16x16_neon;
+    aom_highbd_12_obmc_sub_pixel_variance16x32 = aom_highbd_12_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance16x32 = aom_highbd_12_obmc_sub_pixel_variance16x32_neon;
+    aom_highbd_12_obmc_sub_pixel_variance16x4 = aom_highbd_12_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance16x4 = aom_highbd_12_obmc_sub_pixel_variance16x4_neon;
+    aom_highbd_12_obmc_sub_pixel_variance16x64 = aom_highbd_12_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance16x64 = aom_highbd_12_obmc_sub_pixel_variance16x64_neon;
+    aom_highbd_12_obmc_sub_pixel_variance16x8 = aom_highbd_12_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance16x8 = aom_highbd_12_obmc_sub_pixel_variance16x8_neon;
+    aom_highbd_12_obmc_sub_pixel_variance32x16 = aom_highbd_12_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance32x16 = aom_highbd_12_obmc_sub_pixel_variance32x16_neon;
+    aom_highbd_12_obmc_sub_pixel_variance32x32 = aom_highbd_12_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance32x32 = aom_highbd_12_obmc_sub_pixel_variance32x32_neon;
+    aom_highbd_12_obmc_sub_pixel_variance32x64 = aom_highbd_12_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance32x64 = aom_highbd_12_obmc_sub_pixel_variance32x64_neon;
+    aom_highbd_12_obmc_sub_pixel_variance32x8 = aom_highbd_12_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance32x8 = aom_highbd_12_obmc_sub_pixel_variance32x8_neon;
+    aom_highbd_12_obmc_sub_pixel_variance4x16 = aom_highbd_12_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance4x16 = aom_highbd_12_obmc_sub_pixel_variance4x16_neon;
+    aom_highbd_12_obmc_sub_pixel_variance4x4 = aom_highbd_12_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance4x4 = aom_highbd_12_obmc_sub_pixel_variance4x4_neon;
+    aom_highbd_12_obmc_sub_pixel_variance4x8 = aom_highbd_12_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance4x8 = aom_highbd_12_obmc_sub_pixel_variance4x8_neon;
+    aom_highbd_12_obmc_sub_pixel_variance64x128 = aom_highbd_12_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance64x128 = aom_highbd_12_obmc_sub_pixel_variance64x128_neon;
+    aom_highbd_12_obmc_sub_pixel_variance64x16 = aom_highbd_12_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance64x16 = aom_highbd_12_obmc_sub_pixel_variance64x16_neon;
+    aom_highbd_12_obmc_sub_pixel_variance64x32 = aom_highbd_12_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance64x32 = aom_highbd_12_obmc_sub_pixel_variance64x32_neon;
+    aom_highbd_12_obmc_sub_pixel_variance64x64 = aom_highbd_12_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance64x64 = aom_highbd_12_obmc_sub_pixel_variance64x64_neon;
+    aom_highbd_12_obmc_sub_pixel_variance8x16 = aom_highbd_12_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance8x16 = aom_highbd_12_obmc_sub_pixel_variance8x16_neon;
+    aom_highbd_12_obmc_sub_pixel_variance8x32 = aom_highbd_12_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance8x32 = aom_highbd_12_obmc_sub_pixel_variance8x32_neon;
+    aom_highbd_12_obmc_sub_pixel_variance8x4 = aom_highbd_12_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance8x4 = aom_highbd_12_obmc_sub_pixel_variance8x4_neon;
+    aom_highbd_12_obmc_sub_pixel_variance8x8 = aom_highbd_12_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_sub_pixel_variance8x8 = aom_highbd_12_obmc_sub_pixel_variance8x8_neon;
+    aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_neon;
+    aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_neon;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_neon;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_neon;
+    aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_neon;
+    aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_neon;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_neon;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_neon;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_neon;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_neon;
+    aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_neon;
+    aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_neon;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_neon;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_neon;
+    aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_neon;
+    aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_neon;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_neon;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_neon;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_neon;
+    aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_neon;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_neon;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_neon;
+    aom_highbd_12_sub_pixel_avg_variance128x128 = aom_highbd_12_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance128x128 = aom_highbd_12_sub_pixel_avg_variance128x128_neon;
+    aom_highbd_12_sub_pixel_avg_variance128x64 = aom_highbd_12_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance128x64 = aom_highbd_12_sub_pixel_avg_variance128x64_neon;
+    aom_highbd_12_sub_pixel_avg_variance16x16 = aom_highbd_12_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance16x16 = aom_highbd_12_sub_pixel_avg_variance16x16_neon;
+    aom_highbd_12_sub_pixel_avg_variance16x32 = aom_highbd_12_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance16x32 = aom_highbd_12_sub_pixel_avg_variance16x32_neon;
+    aom_highbd_12_sub_pixel_avg_variance16x4 = aom_highbd_12_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance16x4 = aom_highbd_12_sub_pixel_avg_variance16x4_neon;
+    aom_highbd_12_sub_pixel_avg_variance16x64 = aom_highbd_12_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance16x64 = aom_highbd_12_sub_pixel_avg_variance16x64_neon;
+    aom_highbd_12_sub_pixel_avg_variance16x8 = aom_highbd_12_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance16x8 = aom_highbd_12_sub_pixel_avg_variance16x8_neon;
+    aom_highbd_12_sub_pixel_avg_variance32x16 = aom_highbd_12_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance32x16 = aom_highbd_12_sub_pixel_avg_variance32x16_neon;
+    aom_highbd_12_sub_pixel_avg_variance32x32 = aom_highbd_12_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance32x32 = aom_highbd_12_sub_pixel_avg_variance32x32_neon;
+    aom_highbd_12_sub_pixel_avg_variance32x64 = aom_highbd_12_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance32x64 = aom_highbd_12_sub_pixel_avg_variance32x64_neon;
+    aom_highbd_12_sub_pixel_avg_variance32x8 = aom_highbd_12_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance32x8 = aom_highbd_12_sub_pixel_avg_variance32x8_neon;
+    aom_highbd_12_sub_pixel_avg_variance4x16 = aom_highbd_12_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance4x16 = aom_highbd_12_sub_pixel_avg_variance4x16_neon;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_neon;
+    aom_highbd_12_sub_pixel_avg_variance4x8 = aom_highbd_12_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance4x8 = aom_highbd_12_sub_pixel_avg_variance4x8_neon;
+    aom_highbd_12_sub_pixel_avg_variance64x128 = aom_highbd_12_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance64x128 = aom_highbd_12_sub_pixel_avg_variance64x128_neon;
+    aom_highbd_12_sub_pixel_avg_variance64x16 = aom_highbd_12_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance64x16 = aom_highbd_12_sub_pixel_avg_variance64x16_neon;
+    aom_highbd_12_sub_pixel_avg_variance64x32 = aom_highbd_12_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance64x32 = aom_highbd_12_sub_pixel_avg_variance64x32_neon;
+    aom_highbd_12_sub_pixel_avg_variance64x64 = aom_highbd_12_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance64x64 = aom_highbd_12_sub_pixel_avg_variance64x64_neon;
+    aom_highbd_12_sub_pixel_avg_variance8x16 = aom_highbd_12_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance8x16 = aom_highbd_12_sub_pixel_avg_variance8x16_neon;
+    aom_highbd_12_sub_pixel_avg_variance8x32 = aom_highbd_12_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance8x32 = aom_highbd_12_sub_pixel_avg_variance8x32_neon;
+    aom_highbd_12_sub_pixel_avg_variance8x4 = aom_highbd_12_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance8x4 = aom_highbd_12_sub_pixel_avg_variance8x4_neon;
+    aom_highbd_12_sub_pixel_avg_variance8x8 = aom_highbd_12_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_avg_variance8x8 = aom_highbd_12_sub_pixel_avg_variance8x8_neon;
+    aom_highbd_12_sub_pixel_variance128x128 = aom_highbd_12_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance128x128 = aom_highbd_12_sub_pixel_variance128x128_neon;
+    aom_highbd_12_sub_pixel_variance128x64 = aom_highbd_12_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance128x64 = aom_highbd_12_sub_pixel_variance128x64_neon;
+    aom_highbd_12_sub_pixel_variance16x16 = aom_highbd_12_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance16x16 = aom_highbd_12_sub_pixel_variance16x16_neon;
+    aom_highbd_12_sub_pixel_variance16x32 = aom_highbd_12_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance16x32 = aom_highbd_12_sub_pixel_variance16x32_neon;
+    aom_highbd_12_sub_pixel_variance16x4 = aom_highbd_12_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance16x4 = aom_highbd_12_sub_pixel_variance16x4_neon;
+    aom_highbd_12_sub_pixel_variance16x64 = aom_highbd_12_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance16x64 = aom_highbd_12_sub_pixel_variance16x64_neon;
+    aom_highbd_12_sub_pixel_variance16x8 = aom_highbd_12_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance16x8 = aom_highbd_12_sub_pixel_variance16x8_neon;
+    aom_highbd_12_sub_pixel_variance32x16 = aom_highbd_12_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance32x16 = aom_highbd_12_sub_pixel_variance32x16_neon;
+    aom_highbd_12_sub_pixel_variance32x32 = aom_highbd_12_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance32x32 = aom_highbd_12_sub_pixel_variance32x32_neon;
+    aom_highbd_12_sub_pixel_variance32x64 = aom_highbd_12_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance32x64 = aom_highbd_12_sub_pixel_variance32x64_neon;
+    aom_highbd_12_sub_pixel_variance32x8 = aom_highbd_12_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance32x8 = aom_highbd_12_sub_pixel_variance32x8_neon;
+    aom_highbd_12_sub_pixel_variance4x16 = aom_highbd_12_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance4x16 = aom_highbd_12_sub_pixel_variance4x16_neon;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_neon;
+    aom_highbd_12_sub_pixel_variance4x8 = aom_highbd_12_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance4x8 = aom_highbd_12_sub_pixel_variance4x8_neon;
+    aom_highbd_12_sub_pixel_variance64x128 = aom_highbd_12_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance64x128 = aom_highbd_12_sub_pixel_variance64x128_neon;
+    aom_highbd_12_sub_pixel_variance64x16 = aom_highbd_12_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance64x16 = aom_highbd_12_sub_pixel_variance64x16_neon;
+    aom_highbd_12_sub_pixel_variance64x32 = aom_highbd_12_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance64x32 = aom_highbd_12_sub_pixel_variance64x32_neon;
+    aom_highbd_12_sub_pixel_variance64x64 = aom_highbd_12_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance64x64 = aom_highbd_12_sub_pixel_variance64x64_neon;
+    aom_highbd_12_sub_pixel_variance8x16 = aom_highbd_12_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance8x16 = aom_highbd_12_sub_pixel_variance8x16_neon;
+    aom_highbd_12_sub_pixel_variance8x32 = aom_highbd_12_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance8x32 = aom_highbd_12_sub_pixel_variance8x32_neon;
+    aom_highbd_12_sub_pixel_variance8x4 = aom_highbd_12_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance8x4 = aom_highbd_12_sub_pixel_variance8x4_neon;
+    aom_highbd_12_sub_pixel_variance8x8 = aom_highbd_12_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_sub_pixel_variance8x8 = aom_highbd_12_sub_pixel_variance8x8_neon;
+    aom_highbd_12_variance128x128 = aom_highbd_12_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance128x128 = aom_highbd_12_variance128x128_neon;
+    aom_highbd_12_variance128x64 = aom_highbd_12_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance128x64 = aom_highbd_12_variance128x64_neon;
+    aom_highbd_12_variance16x16 = aom_highbd_12_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance16x16 = aom_highbd_12_variance16x16_neon;
+    aom_highbd_12_variance16x32 = aom_highbd_12_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance16x32 = aom_highbd_12_variance16x32_neon;
+    aom_highbd_12_variance16x4 = aom_highbd_12_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance16x4 = aom_highbd_12_variance16x4_neon;
+    aom_highbd_12_variance16x64 = aom_highbd_12_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance16x64 = aom_highbd_12_variance16x64_neon;
+    aom_highbd_12_variance16x8 = aom_highbd_12_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance16x8 = aom_highbd_12_variance16x8_neon;
+    aom_highbd_12_variance32x16 = aom_highbd_12_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance32x16 = aom_highbd_12_variance32x16_neon;
+    aom_highbd_12_variance32x32 = aom_highbd_12_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance32x32 = aom_highbd_12_variance32x32_neon;
+    aom_highbd_12_variance32x64 = aom_highbd_12_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance32x64 = aom_highbd_12_variance32x64_neon;
+    aom_highbd_12_variance32x8 = aom_highbd_12_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance32x8 = aom_highbd_12_variance32x8_neon;
+    aom_highbd_12_variance4x16 = aom_highbd_12_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance4x16 = aom_highbd_12_variance4x16_neon;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_neon;
+    aom_highbd_12_variance4x8 = aom_highbd_12_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance4x8 = aom_highbd_12_variance4x8_neon;
+    aom_highbd_12_variance64x128 = aom_highbd_12_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance64x128 = aom_highbd_12_variance64x128_neon;
+    aom_highbd_12_variance64x16 = aom_highbd_12_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance64x16 = aom_highbd_12_variance64x16_neon;
+    aom_highbd_12_variance64x32 = aom_highbd_12_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance64x32 = aom_highbd_12_variance64x32_neon;
+    aom_highbd_12_variance64x64 = aom_highbd_12_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance64x64 = aom_highbd_12_variance64x64_neon;
+    aom_highbd_12_variance8x16 = aom_highbd_12_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance8x16 = aom_highbd_12_variance8x16_neon;
+    aom_highbd_12_variance8x32 = aom_highbd_12_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance8x32 = aom_highbd_12_variance8x32_neon;
+    aom_highbd_12_variance8x4 = aom_highbd_12_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance8x4 = aom_highbd_12_variance8x4_neon;
+    aom_highbd_12_variance8x8 = aom_highbd_12_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_12_variance8x8 = aom_highbd_12_variance8x8_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_neon;
+    aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 = aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_neon;
+    aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_neon;
+    aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_neon;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_neon;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_neon;
+    aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_neon;
+    aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_neon;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_neon;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_neon;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_neon;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_neon;
+    aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_neon;
+    aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_neon;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_neon;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_neon;
+    aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_neon;
+    aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_neon;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_neon;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_neon;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_neon;
+    aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_neon;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_neon;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_neon;
+    aom_highbd_8_mse16x16 = aom_highbd_8_mse16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_mse16x16 = aom_highbd_8_mse16x16_neon;
+    aom_highbd_8_mse16x8 = aom_highbd_8_mse16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_mse16x8 = aom_highbd_8_mse16x8_neon;
+    aom_highbd_8_mse8x16 = aom_highbd_8_mse8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_mse8x16 = aom_highbd_8_mse8x16_neon;
+    aom_highbd_8_mse8x8 = aom_highbd_8_mse8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_mse8x8 = aom_highbd_8_mse8x8_neon;
+    aom_highbd_8_obmc_sub_pixel_variance128x128 = aom_highbd_8_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance128x128 = aom_highbd_8_obmc_sub_pixel_variance128x128_neon;
+    aom_highbd_8_obmc_sub_pixel_variance128x64 = aom_highbd_8_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance128x64 = aom_highbd_8_obmc_sub_pixel_variance128x64_neon;
+    aom_highbd_8_obmc_sub_pixel_variance16x16 = aom_highbd_8_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance16x16 = aom_highbd_8_obmc_sub_pixel_variance16x16_neon;
+    aom_highbd_8_obmc_sub_pixel_variance16x32 = aom_highbd_8_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance16x32 = aom_highbd_8_obmc_sub_pixel_variance16x32_neon;
+    aom_highbd_8_obmc_sub_pixel_variance16x4 = aom_highbd_8_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance16x4 = aom_highbd_8_obmc_sub_pixel_variance16x4_neon;
+    aom_highbd_8_obmc_sub_pixel_variance16x64 = aom_highbd_8_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance16x64 = aom_highbd_8_obmc_sub_pixel_variance16x64_neon;
+    aom_highbd_8_obmc_sub_pixel_variance16x8 = aom_highbd_8_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance16x8 = aom_highbd_8_obmc_sub_pixel_variance16x8_neon;
+    aom_highbd_8_obmc_sub_pixel_variance32x16 = aom_highbd_8_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance32x16 = aom_highbd_8_obmc_sub_pixel_variance32x16_neon;
+    aom_highbd_8_obmc_sub_pixel_variance32x32 = aom_highbd_8_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance32x32 = aom_highbd_8_obmc_sub_pixel_variance32x32_neon;
+    aom_highbd_8_obmc_sub_pixel_variance32x64 = aom_highbd_8_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance32x64 = aom_highbd_8_obmc_sub_pixel_variance32x64_neon;
+    aom_highbd_8_obmc_sub_pixel_variance32x8 = aom_highbd_8_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance32x8 = aom_highbd_8_obmc_sub_pixel_variance32x8_neon;
+    aom_highbd_8_obmc_sub_pixel_variance4x16 = aom_highbd_8_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance4x16 = aom_highbd_8_obmc_sub_pixel_variance4x16_neon;
+    aom_highbd_8_obmc_sub_pixel_variance4x4 = aom_highbd_8_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance4x4 = aom_highbd_8_obmc_sub_pixel_variance4x4_neon;
+    aom_highbd_8_obmc_sub_pixel_variance4x8 = aom_highbd_8_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance4x8 = aom_highbd_8_obmc_sub_pixel_variance4x8_neon;
+    aom_highbd_8_obmc_sub_pixel_variance64x128 = aom_highbd_8_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance64x128 = aom_highbd_8_obmc_sub_pixel_variance64x128_neon;
+    aom_highbd_8_obmc_sub_pixel_variance64x16 = aom_highbd_8_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance64x16 = aom_highbd_8_obmc_sub_pixel_variance64x16_neon;
+    aom_highbd_8_obmc_sub_pixel_variance64x32 = aom_highbd_8_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance64x32 = aom_highbd_8_obmc_sub_pixel_variance64x32_neon;
+    aom_highbd_8_obmc_sub_pixel_variance64x64 = aom_highbd_8_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance64x64 = aom_highbd_8_obmc_sub_pixel_variance64x64_neon;
+    aom_highbd_8_obmc_sub_pixel_variance8x16 = aom_highbd_8_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance8x16 = aom_highbd_8_obmc_sub_pixel_variance8x16_neon;
+    aom_highbd_8_obmc_sub_pixel_variance8x32 = aom_highbd_8_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance8x32 = aom_highbd_8_obmc_sub_pixel_variance8x32_neon;
+    aom_highbd_8_obmc_sub_pixel_variance8x4 = aom_highbd_8_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance8x4 = aom_highbd_8_obmc_sub_pixel_variance8x4_neon;
+    aom_highbd_8_obmc_sub_pixel_variance8x8 = aom_highbd_8_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_sub_pixel_variance8x8 = aom_highbd_8_obmc_sub_pixel_variance8x8_neon;
+    aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_neon;
+    aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_neon;
+    aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_neon;
+    aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_neon;
+    aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_neon;
+    aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_neon;
+    aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_neon;
+    aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_neon;
+    aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_neon;
+    aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_neon;
+    aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_neon;
+    aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_neon;
+    aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_neon;
+    aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_neon;
+    aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_neon;
+    aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_neon;
+    aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_neon;
+    aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_neon;
+    aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_neon;
+    aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_neon;
+    aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_neon;
+    aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_neon;
+    aom_highbd_8_sub_pixel_avg_variance128x128 = aom_highbd_8_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance128x128 = aom_highbd_8_sub_pixel_avg_variance128x128_neon;
+    aom_highbd_8_sub_pixel_avg_variance128x64 = aom_highbd_8_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance128x64 = aom_highbd_8_sub_pixel_avg_variance128x64_neon;
+    aom_highbd_8_sub_pixel_avg_variance16x16 = aom_highbd_8_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance16x16 = aom_highbd_8_sub_pixel_avg_variance16x16_neon;
+    aom_highbd_8_sub_pixel_avg_variance16x32 = aom_highbd_8_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance16x32 = aom_highbd_8_sub_pixel_avg_variance16x32_neon;
+    aom_highbd_8_sub_pixel_avg_variance16x4 = aom_highbd_8_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance16x4 = aom_highbd_8_sub_pixel_avg_variance16x4_neon;
+    aom_highbd_8_sub_pixel_avg_variance16x64 = aom_highbd_8_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance16x64 = aom_highbd_8_sub_pixel_avg_variance16x64_neon;
+    aom_highbd_8_sub_pixel_avg_variance16x8 = aom_highbd_8_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance16x8 = aom_highbd_8_sub_pixel_avg_variance16x8_neon;
+    aom_highbd_8_sub_pixel_avg_variance32x16 = aom_highbd_8_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance32x16 = aom_highbd_8_sub_pixel_avg_variance32x16_neon;
+    aom_highbd_8_sub_pixel_avg_variance32x32 = aom_highbd_8_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance32x32 = aom_highbd_8_sub_pixel_avg_variance32x32_neon;
+    aom_highbd_8_sub_pixel_avg_variance32x64 = aom_highbd_8_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance32x64 = aom_highbd_8_sub_pixel_avg_variance32x64_neon;
+    aom_highbd_8_sub_pixel_avg_variance32x8 = aom_highbd_8_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance32x8 = aom_highbd_8_sub_pixel_avg_variance32x8_neon;
+    aom_highbd_8_sub_pixel_avg_variance4x16 = aom_highbd_8_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance4x16 = aom_highbd_8_sub_pixel_avg_variance4x16_neon;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_neon;
+    aom_highbd_8_sub_pixel_avg_variance4x8 = aom_highbd_8_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance4x8 = aom_highbd_8_sub_pixel_avg_variance4x8_neon;
+    aom_highbd_8_sub_pixel_avg_variance64x128 = aom_highbd_8_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance64x128 = aom_highbd_8_sub_pixel_avg_variance64x128_neon;
+    aom_highbd_8_sub_pixel_avg_variance64x16 = aom_highbd_8_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance64x16 = aom_highbd_8_sub_pixel_avg_variance64x16_neon;
+    aom_highbd_8_sub_pixel_avg_variance64x32 = aom_highbd_8_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance64x32 = aom_highbd_8_sub_pixel_avg_variance64x32_neon;
+    aom_highbd_8_sub_pixel_avg_variance64x64 = aom_highbd_8_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance64x64 = aom_highbd_8_sub_pixel_avg_variance64x64_neon;
+    aom_highbd_8_sub_pixel_avg_variance8x16 = aom_highbd_8_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance8x16 = aom_highbd_8_sub_pixel_avg_variance8x16_neon;
+    aom_highbd_8_sub_pixel_avg_variance8x32 = aom_highbd_8_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance8x32 = aom_highbd_8_sub_pixel_avg_variance8x32_neon;
+    aom_highbd_8_sub_pixel_avg_variance8x4 = aom_highbd_8_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance8x4 = aom_highbd_8_sub_pixel_avg_variance8x4_neon;
+    aom_highbd_8_sub_pixel_avg_variance8x8 = aom_highbd_8_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_avg_variance8x8 = aom_highbd_8_sub_pixel_avg_variance8x8_neon;
+    aom_highbd_8_sub_pixel_variance128x128 = aom_highbd_8_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance128x128 = aom_highbd_8_sub_pixel_variance128x128_neon;
+    aom_highbd_8_sub_pixel_variance128x64 = aom_highbd_8_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance128x64 = aom_highbd_8_sub_pixel_variance128x64_neon;
+    aom_highbd_8_sub_pixel_variance16x16 = aom_highbd_8_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance16x16 = aom_highbd_8_sub_pixel_variance16x16_neon;
+    aom_highbd_8_sub_pixel_variance16x32 = aom_highbd_8_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance16x32 = aom_highbd_8_sub_pixel_variance16x32_neon;
+    aom_highbd_8_sub_pixel_variance16x4 = aom_highbd_8_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance16x4 = aom_highbd_8_sub_pixel_variance16x4_neon;
+    aom_highbd_8_sub_pixel_variance16x64 = aom_highbd_8_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance16x64 = aom_highbd_8_sub_pixel_variance16x64_neon;
+    aom_highbd_8_sub_pixel_variance16x8 = aom_highbd_8_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance16x8 = aom_highbd_8_sub_pixel_variance16x8_neon;
+    aom_highbd_8_sub_pixel_variance32x16 = aom_highbd_8_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance32x16 = aom_highbd_8_sub_pixel_variance32x16_neon;
+    aom_highbd_8_sub_pixel_variance32x32 = aom_highbd_8_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance32x32 = aom_highbd_8_sub_pixel_variance32x32_neon;
+    aom_highbd_8_sub_pixel_variance32x64 = aom_highbd_8_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance32x64 = aom_highbd_8_sub_pixel_variance32x64_neon;
+    aom_highbd_8_sub_pixel_variance32x8 = aom_highbd_8_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance32x8 = aom_highbd_8_sub_pixel_variance32x8_neon;
+    aom_highbd_8_sub_pixel_variance4x16 = aom_highbd_8_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance4x16 = aom_highbd_8_sub_pixel_variance4x16_neon;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_neon;
+    aom_highbd_8_sub_pixel_variance4x8 = aom_highbd_8_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance4x8 = aom_highbd_8_sub_pixel_variance4x8_neon;
+    aom_highbd_8_sub_pixel_variance64x128 = aom_highbd_8_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance64x128 = aom_highbd_8_sub_pixel_variance64x128_neon;
+    aom_highbd_8_sub_pixel_variance64x16 = aom_highbd_8_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance64x16 = aom_highbd_8_sub_pixel_variance64x16_neon;
+    aom_highbd_8_sub_pixel_variance64x32 = aom_highbd_8_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance64x32 = aom_highbd_8_sub_pixel_variance64x32_neon;
+    aom_highbd_8_sub_pixel_variance64x64 = aom_highbd_8_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance64x64 = aom_highbd_8_sub_pixel_variance64x64_neon;
+    aom_highbd_8_sub_pixel_variance8x16 = aom_highbd_8_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance8x16 = aom_highbd_8_sub_pixel_variance8x16_neon;
+    aom_highbd_8_sub_pixel_variance8x32 = aom_highbd_8_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance8x32 = aom_highbd_8_sub_pixel_variance8x32_neon;
+    aom_highbd_8_sub_pixel_variance8x4 = aom_highbd_8_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance8x4 = aom_highbd_8_sub_pixel_variance8x4_neon;
+    aom_highbd_8_sub_pixel_variance8x8 = aom_highbd_8_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_sub_pixel_variance8x8 = aom_highbd_8_sub_pixel_variance8x8_neon;
+    aom_highbd_8_variance128x128 = aom_highbd_8_variance128x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance128x128 = aom_highbd_8_variance128x128_neon;
+    aom_highbd_8_variance128x64 = aom_highbd_8_variance128x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance128x64 = aom_highbd_8_variance128x64_neon;
+    aom_highbd_8_variance16x16 = aom_highbd_8_variance16x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance16x16 = aom_highbd_8_variance16x16_neon;
+    aom_highbd_8_variance16x32 = aom_highbd_8_variance16x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance16x32 = aom_highbd_8_variance16x32_neon;
+    aom_highbd_8_variance16x4 = aom_highbd_8_variance16x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance16x4 = aom_highbd_8_variance16x4_neon;
+    aom_highbd_8_variance16x64 = aom_highbd_8_variance16x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance16x64 = aom_highbd_8_variance16x64_neon;
+    aom_highbd_8_variance16x8 = aom_highbd_8_variance16x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance16x8 = aom_highbd_8_variance16x8_neon;
+    aom_highbd_8_variance32x16 = aom_highbd_8_variance32x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance32x16 = aom_highbd_8_variance32x16_neon;
+    aom_highbd_8_variance32x32 = aom_highbd_8_variance32x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance32x32 = aom_highbd_8_variance32x32_neon;
+    aom_highbd_8_variance32x64 = aom_highbd_8_variance32x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance32x64 = aom_highbd_8_variance32x64_neon;
+    aom_highbd_8_variance32x8 = aom_highbd_8_variance32x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance32x8 = aom_highbd_8_variance32x8_neon;
+    aom_highbd_8_variance4x16 = aom_highbd_8_variance4x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance4x16 = aom_highbd_8_variance4x16_neon;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_neon;
+    aom_highbd_8_variance4x8 = aom_highbd_8_variance4x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance4x8 = aom_highbd_8_variance4x8_neon;
+    aom_highbd_8_variance64x128 = aom_highbd_8_variance64x128_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance64x128 = aom_highbd_8_variance64x128_neon;
+    aom_highbd_8_variance64x16 = aom_highbd_8_variance64x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance64x16 = aom_highbd_8_variance64x16_neon;
+    aom_highbd_8_variance64x32 = aom_highbd_8_variance64x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance64x32 = aom_highbd_8_variance64x32_neon;
+    aom_highbd_8_variance64x64 = aom_highbd_8_variance64x64_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance64x64 = aom_highbd_8_variance64x64_neon;
+    aom_highbd_8_variance8x16 = aom_highbd_8_variance8x16_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance8x16 = aom_highbd_8_variance8x16_neon;
+    aom_highbd_8_variance8x32 = aom_highbd_8_variance8x32_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance8x32 = aom_highbd_8_variance8x32_neon;
+    aom_highbd_8_variance8x4 = aom_highbd_8_variance8x4_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance8x4 = aom_highbd_8_variance8x4_neon;
+    aom_highbd_8_variance8x8 = aom_highbd_8_variance8x8_c;
+    if (flags & HAS_NEON) aom_highbd_8_variance8x8 = aom_highbd_8_variance8x8_neon;
+    aom_highbd_avg_4x4 = aom_highbd_avg_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_avg_4x4 = aom_highbd_avg_4x4_neon;
+    aom_highbd_avg_8x8 = aom_highbd_avg_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_avg_8x8 = aom_highbd_avg_8x8_neon;
+    aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_c;
+    if (flags & HAS_NEON) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_neon;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_NEON) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_neon;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_NEON) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_neon;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_NEON) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_neon;
+    aom_highbd_comp_avg_pred = aom_highbd_comp_avg_pred_c;
+    if (flags & HAS_NEON) aom_highbd_comp_avg_pred = aom_highbd_comp_avg_pred_neon;
+    aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_c;
+    if (flags & HAS_NEON) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_neon;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_c;
+    if (flags & HAS_NEON) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_neon;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_c;
+    if (flags & HAS_NEON) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_neon;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_c;
+    if (flags & HAS_NEON) aom_highbd_convolve_copy = aom_highbd_convolve_copy_neon;
+    aom_highbd_dc_128_predictor_16x16 = aom_highbd_dc_128_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_16x16 = aom_highbd_dc_128_predictor_16x16_neon;
+    aom_highbd_dc_128_predictor_16x32 = aom_highbd_dc_128_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_16x32 = aom_highbd_dc_128_predictor_16x32_neon;
+    aom_highbd_dc_128_predictor_16x4 = aom_highbd_dc_128_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_16x4 = aom_highbd_dc_128_predictor_16x4_neon;
+    aom_highbd_dc_128_predictor_16x64 = aom_highbd_dc_128_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_16x64 = aom_highbd_dc_128_predictor_16x64_neon;
+    aom_highbd_dc_128_predictor_16x8 = aom_highbd_dc_128_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_16x8 = aom_highbd_dc_128_predictor_16x8_neon;
+    aom_highbd_dc_128_predictor_32x16 = aom_highbd_dc_128_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_32x16 = aom_highbd_dc_128_predictor_32x16_neon;
+    aom_highbd_dc_128_predictor_32x32 = aom_highbd_dc_128_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_32x32 = aom_highbd_dc_128_predictor_32x32_neon;
+    aom_highbd_dc_128_predictor_32x64 = aom_highbd_dc_128_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_32x64 = aom_highbd_dc_128_predictor_32x64_neon;
+    aom_highbd_dc_128_predictor_32x8 = aom_highbd_dc_128_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_32x8 = aom_highbd_dc_128_predictor_32x8_neon;
+    aom_highbd_dc_128_predictor_4x16 = aom_highbd_dc_128_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_4x16 = aom_highbd_dc_128_predictor_4x16_neon;
+    aom_highbd_dc_128_predictor_4x4 = aom_highbd_dc_128_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_4x4 = aom_highbd_dc_128_predictor_4x4_neon;
+    aom_highbd_dc_128_predictor_4x8 = aom_highbd_dc_128_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_4x8 = aom_highbd_dc_128_predictor_4x8_neon;
+    aom_highbd_dc_128_predictor_64x16 = aom_highbd_dc_128_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_64x16 = aom_highbd_dc_128_predictor_64x16_neon;
+    aom_highbd_dc_128_predictor_64x32 = aom_highbd_dc_128_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_64x32 = aom_highbd_dc_128_predictor_64x32_neon;
+    aom_highbd_dc_128_predictor_64x64 = aom_highbd_dc_128_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_64x64 = aom_highbd_dc_128_predictor_64x64_neon;
+    aom_highbd_dc_128_predictor_8x16 = aom_highbd_dc_128_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_8x16 = aom_highbd_dc_128_predictor_8x16_neon;
+    aom_highbd_dc_128_predictor_8x32 = aom_highbd_dc_128_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_8x32 = aom_highbd_dc_128_predictor_8x32_neon;
+    aom_highbd_dc_128_predictor_8x4 = aom_highbd_dc_128_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_8x4 = aom_highbd_dc_128_predictor_8x4_neon;
+    aom_highbd_dc_128_predictor_8x8 = aom_highbd_dc_128_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_128_predictor_8x8 = aom_highbd_dc_128_predictor_8x8_neon;
+    aom_highbd_dc_left_predictor_16x16 = aom_highbd_dc_left_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_16x16 = aom_highbd_dc_left_predictor_16x16_neon;
+    aom_highbd_dc_left_predictor_16x32 = aom_highbd_dc_left_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_16x32 = aom_highbd_dc_left_predictor_16x32_neon;
+    aom_highbd_dc_left_predictor_16x4 = aom_highbd_dc_left_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_16x4 = aom_highbd_dc_left_predictor_16x4_neon;
+    aom_highbd_dc_left_predictor_16x64 = aom_highbd_dc_left_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_16x64 = aom_highbd_dc_left_predictor_16x64_neon;
+    aom_highbd_dc_left_predictor_16x8 = aom_highbd_dc_left_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_16x8 = aom_highbd_dc_left_predictor_16x8_neon;
+    aom_highbd_dc_left_predictor_32x16 = aom_highbd_dc_left_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_32x16 = aom_highbd_dc_left_predictor_32x16_neon;
+    aom_highbd_dc_left_predictor_32x32 = aom_highbd_dc_left_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_32x32 = aom_highbd_dc_left_predictor_32x32_neon;
+    aom_highbd_dc_left_predictor_32x64 = aom_highbd_dc_left_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_32x64 = aom_highbd_dc_left_predictor_32x64_neon;
+    aom_highbd_dc_left_predictor_32x8 = aom_highbd_dc_left_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_32x8 = aom_highbd_dc_left_predictor_32x8_neon;
+    aom_highbd_dc_left_predictor_4x16 = aom_highbd_dc_left_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_4x16 = aom_highbd_dc_left_predictor_4x16_neon;
+    aom_highbd_dc_left_predictor_4x4 = aom_highbd_dc_left_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_4x4 = aom_highbd_dc_left_predictor_4x4_neon;
+    aom_highbd_dc_left_predictor_4x8 = aom_highbd_dc_left_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_4x8 = aom_highbd_dc_left_predictor_4x8_neon;
+    aom_highbd_dc_left_predictor_64x16 = aom_highbd_dc_left_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_64x16 = aom_highbd_dc_left_predictor_64x16_neon;
+    aom_highbd_dc_left_predictor_64x32 = aom_highbd_dc_left_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_64x32 = aom_highbd_dc_left_predictor_64x32_neon;
+    aom_highbd_dc_left_predictor_64x64 = aom_highbd_dc_left_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_64x64 = aom_highbd_dc_left_predictor_64x64_neon;
+    aom_highbd_dc_left_predictor_8x16 = aom_highbd_dc_left_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_8x16 = aom_highbd_dc_left_predictor_8x16_neon;
+    aom_highbd_dc_left_predictor_8x32 = aom_highbd_dc_left_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_8x32 = aom_highbd_dc_left_predictor_8x32_neon;
+    aom_highbd_dc_left_predictor_8x4 = aom_highbd_dc_left_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_8x4 = aom_highbd_dc_left_predictor_8x4_neon;
+    aom_highbd_dc_left_predictor_8x8 = aom_highbd_dc_left_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_left_predictor_8x8 = aom_highbd_dc_left_predictor_8x8_neon;
+    aom_highbd_dc_predictor_16x16 = aom_highbd_dc_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_16x16 = aom_highbd_dc_predictor_16x16_neon;
+    aom_highbd_dc_predictor_16x32 = aom_highbd_dc_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_16x32 = aom_highbd_dc_predictor_16x32_neon;
+    aom_highbd_dc_predictor_16x4 = aom_highbd_dc_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_16x4 = aom_highbd_dc_predictor_16x4_neon;
+    aom_highbd_dc_predictor_16x64 = aom_highbd_dc_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_16x64 = aom_highbd_dc_predictor_16x64_neon;
+    aom_highbd_dc_predictor_16x8 = aom_highbd_dc_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_16x8 = aom_highbd_dc_predictor_16x8_neon;
+    aom_highbd_dc_predictor_32x16 = aom_highbd_dc_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_32x16 = aom_highbd_dc_predictor_32x16_neon;
+    aom_highbd_dc_predictor_32x32 = aom_highbd_dc_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_32x32 = aom_highbd_dc_predictor_32x32_neon;
+    aom_highbd_dc_predictor_32x64 = aom_highbd_dc_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_32x64 = aom_highbd_dc_predictor_32x64_neon;
+    aom_highbd_dc_predictor_32x8 = aom_highbd_dc_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_32x8 = aom_highbd_dc_predictor_32x8_neon;
+    aom_highbd_dc_predictor_4x16 = aom_highbd_dc_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_4x16 = aom_highbd_dc_predictor_4x16_neon;
+    aom_highbd_dc_predictor_4x4 = aom_highbd_dc_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_4x4 = aom_highbd_dc_predictor_4x4_neon;
+    aom_highbd_dc_predictor_4x8 = aom_highbd_dc_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_4x8 = aom_highbd_dc_predictor_4x8_neon;
+    aom_highbd_dc_predictor_64x16 = aom_highbd_dc_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_64x16 = aom_highbd_dc_predictor_64x16_neon;
+    aom_highbd_dc_predictor_64x32 = aom_highbd_dc_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_64x32 = aom_highbd_dc_predictor_64x32_neon;
+    aom_highbd_dc_predictor_64x64 = aom_highbd_dc_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_64x64 = aom_highbd_dc_predictor_64x64_neon;
+    aom_highbd_dc_predictor_8x16 = aom_highbd_dc_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_8x16 = aom_highbd_dc_predictor_8x16_neon;
+    aom_highbd_dc_predictor_8x32 = aom_highbd_dc_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_8x32 = aom_highbd_dc_predictor_8x32_neon;
+    aom_highbd_dc_predictor_8x4 = aom_highbd_dc_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_8x4 = aom_highbd_dc_predictor_8x4_neon;
+    aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_neon;
+    aom_highbd_dc_top_predictor_16x16 = aom_highbd_dc_top_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_16x16 = aom_highbd_dc_top_predictor_16x16_neon;
+    aom_highbd_dc_top_predictor_16x32 = aom_highbd_dc_top_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_16x32 = aom_highbd_dc_top_predictor_16x32_neon;
+    aom_highbd_dc_top_predictor_16x4 = aom_highbd_dc_top_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_16x4 = aom_highbd_dc_top_predictor_16x4_neon;
+    aom_highbd_dc_top_predictor_16x64 = aom_highbd_dc_top_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_16x64 = aom_highbd_dc_top_predictor_16x64_neon;
+    aom_highbd_dc_top_predictor_16x8 = aom_highbd_dc_top_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_16x8 = aom_highbd_dc_top_predictor_16x8_neon;
+    aom_highbd_dc_top_predictor_32x16 = aom_highbd_dc_top_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_32x16 = aom_highbd_dc_top_predictor_32x16_neon;
+    aom_highbd_dc_top_predictor_32x32 = aom_highbd_dc_top_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_32x32 = aom_highbd_dc_top_predictor_32x32_neon;
+    aom_highbd_dc_top_predictor_32x64 = aom_highbd_dc_top_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_32x64 = aom_highbd_dc_top_predictor_32x64_neon;
+    aom_highbd_dc_top_predictor_32x8 = aom_highbd_dc_top_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_32x8 = aom_highbd_dc_top_predictor_32x8_neon;
+    aom_highbd_dc_top_predictor_4x16 = aom_highbd_dc_top_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_4x16 = aom_highbd_dc_top_predictor_4x16_neon;
+    aom_highbd_dc_top_predictor_4x4 = aom_highbd_dc_top_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_4x4 = aom_highbd_dc_top_predictor_4x4_neon;
+    aom_highbd_dc_top_predictor_4x8 = aom_highbd_dc_top_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_4x8 = aom_highbd_dc_top_predictor_4x8_neon;
+    aom_highbd_dc_top_predictor_64x16 = aom_highbd_dc_top_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_64x16 = aom_highbd_dc_top_predictor_64x16_neon;
+    aom_highbd_dc_top_predictor_64x32 = aom_highbd_dc_top_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_64x32 = aom_highbd_dc_top_predictor_64x32_neon;
+    aom_highbd_dc_top_predictor_64x64 = aom_highbd_dc_top_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_64x64 = aom_highbd_dc_top_predictor_64x64_neon;
+    aom_highbd_dc_top_predictor_8x16 = aom_highbd_dc_top_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_8x16 = aom_highbd_dc_top_predictor_8x16_neon;
+    aom_highbd_dc_top_predictor_8x32 = aom_highbd_dc_top_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_8x32 = aom_highbd_dc_top_predictor_8x32_neon;
+    aom_highbd_dc_top_predictor_8x4 = aom_highbd_dc_top_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_8x4 = aom_highbd_dc_top_predictor_8x4_neon;
+    aom_highbd_dc_top_predictor_8x8 = aom_highbd_dc_top_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_dc_top_predictor_8x8 = aom_highbd_dc_top_predictor_8x8_neon;
+    aom_highbd_dist_wtd_comp_avg_pred = aom_highbd_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_NEON) aom_highbd_dist_wtd_comp_avg_pred = aom_highbd_dist_wtd_comp_avg_pred_neon;
+    aom_highbd_h_predictor_16x16 = aom_highbd_h_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_16x16 = aom_highbd_h_predictor_16x16_neon;
+    aom_highbd_h_predictor_16x32 = aom_highbd_h_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_16x32 = aom_highbd_h_predictor_16x32_neon;
+    aom_highbd_h_predictor_16x4 = aom_highbd_h_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_16x4 = aom_highbd_h_predictor_16x4_neon;
+    aom_highbd_h_predictor_16x64 = aom_highbd_h_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_16x64 = aom_highbd_h_predictor_16x64_neon;
+    aom_highbd_h_predictor_16x8 = aom_highbd_h_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_16x8 = aom_highbd_h_predictor_16x8_neon;
+    aom_highbd_h_predictor_32x16 = aom_highbd_h_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_32x16 = aom_highbd_h_predictor_32x16_neon;
+    aom_highbd_h_predictor_32x32 = aom_highbd_h_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_32x32 = aom_highbd_h_predictor_32x32_neon;
+    aom_highbd_h_predictor_32x64 = aom_highbd_h_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_32x64 = aom_highbd_h_predictor_32x64_neon;
+    aom_highbd_h_predictor_32x8 = aom_highbd_h_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_32x8 = aom_highbd_h_predictor_32x8_neon;
+    aom_highbd_h_predictor_4x16 = aom_highbd_h_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_4x16 = aom_highbd_h_predictor_4x16_neon;
+    aom_highbd_h_predictor_4x4 = aom_highbd_h_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_4x4 = aom_highbd_h_predictor_4x4_neon;
+    aom_highbd_h_predictor_4x8 = aom_highbd_h_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_4x8 = aom_highbd_h_predictor_4x8_neon;
+    aom_highbd_h_predictor_64x16 = aom_highbd_h_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_64x16 = aom_highbd_h_predictor_64x16_neon;
+    aom_highbd_h_predictor_64x32 = aom_highbd_h_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_64x32 = aom_highbd_h_predictor_64x32_neon;
+    aom_highbd_h_predictor_64x64 = aom_highbd_h_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_64x64 = aom_highbd_h_predictor_64x64_neon;
+    aom_highbd_h_predictor_8x16 = aom_highbd_h_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_8x16 = aom_highbd_h_predictor_8x16_neon;
+    aom_highbd_h_predictor_8x32 = aom_highbd_h_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_8x32 = aom_highbd_h_predictor_8x32_neon;
+    aom_highbd_h_predictor_8x4 = aom_highbd_h_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_8x4 = aom_highbd_h_predictor_8x4_neon;
+    aom_highbd_h_predictor_8x8 = aom_highbd_h_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_h_predictor_8x8 = aom_highbd_h_predictor_8x8_neon;
+    aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_neon;
+    aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_neon;
+    aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_neon;
+    aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_neon;
+    aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_neon;
+    aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_neon;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_neon;
+    aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_neon;
+    aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_neon;
+    aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_neon;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_neon;
+    aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_neon;
+    aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_neon;
+    aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_neon;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_neon;
+    aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_neon;
+    aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_neon;
+    aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_neon;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_c;
+    if (flags & HAS_NEON) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_neon;
+    aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_neon;
+    aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_neon;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_neon;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_neon;
+    aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_neon;
+    aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_neon;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_neon;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_neon;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_neon;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_neon;
+    aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_neon;
+    aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_neon;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_neon;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_neon;
+    aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_neon;
+    aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_neon;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_neon;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_neon;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_neon;
+    aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_neon;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_neon;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_NEON) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_neon;
+    aom_highbd_minmax_8x8 = aom_highbd_minmax_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_minmax_8x8 = aom_highbd_minmax_8x8_neon;
+    aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_neon;
+    aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_neon;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_neon;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_neon;
+    aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_neon;
+    aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_neon;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_neon;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_neon;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_neon;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_neon;
+    aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_neon;
+    aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_neon;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_neon;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_neon;
+    aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_neon;
+    aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_neon;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_neon;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_neon;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_neon;
+    aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_neon;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_neon;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_NEON) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_neon;
+    aom_highbd_paeth_predictor_16x16 = aom_highbd_paeth_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x16 = aom_highbd_paeth_predictor_16x16_neon;
+    aom_highbd_paeth_predictor_16x32 = aom_highbd_paeth_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x32 = aom_highbd_paeth_predictor_16x32_neon;
+    aom_highbd_paeth_predictor_16x4 = aom_highbd_paeth_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x4 = aom_highbd_paeth_predictor_16x4_neon;
+    aom_highbd_paeth_predictor_16x64 = aom_highbd_paeth_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x64 = aom_highbd_paeth_predictor_16x64_neon;
+    aom_highbd_paeth_predictor_16x8 = aom_highbd_paeth_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x8 = aom_highbd_paeth_predictor_16x8_neon;
+    aom_highbd_paeth_predictor_32x16 = aom_highbd_paeth_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x16 = aom_highbd_paeth_predictor_32x16_neon;
+    aom_highbd_paeth_predictor_32x32 = aom_highbd_paeth_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x32 = aom_highbd_paeth_predictor_32x32_neon;
+    aom_highbd_paeth_predictor_32x64 = aom_highbd_paeth_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x64 = aom_highbd_paeth_predictor_32x64_neon;
+    aom_highbd_paeth_predictor_32x8 = aom_highbd_paeth_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x8 = aom_highbd_paeth_predictor_32x8_neon;
+    aom_highbd_paeth_predictor_4x16 = aom_highbd_paeth_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_4x16 = aom_highbd_paeth_predictor_4x16_neon;
+    aom_highbd_paeth_predictor_4x4 = aom_highbd_paeth_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_4x4 = aom_highbd_paeth_predictor_4x4_neon;
+    aom_highbd_paeth_predictor_4x8 = aom_highbd_paeth_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_4x8 = aom_highbd_paeth_predictor_4x8_neon;
+    aom_highbd_paeth_predictor_64x16 = aom_highbd_paeth_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_64x16 = aom_highbd_paeth_predictor_64x16_neon;
+    aom_highbd_paeth_predictor_64x32 = aom_highbd_paeth_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_64x32 = aom_highbd_paeth_predictor_64x32_neon;
+    aom_highbd_paeth_predictor_64x64 = aom_highbd_paeth_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_64x64 = aom_highbd_paeth_predictor_64x64_neon;
+    aom_highbd_paeth_predictor_8x16 = aom_highbd_paeth_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x16 = aom_highbd_paeth_predictor_8x16_neon;
+    aom_highbd_paeth_predictor_8x32 = aom_highbd_paeth_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x32 = aom_highbd_paeth_predictor_8x32_neon;
+    aom_highbd_paeth_predictor_8x4 = aom_highbd_paeth_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x4 = aom_highbd_paeth_predictor_8x4_neon;
+    aom_highbd_paeth_predictor_8x8 = aom_highbd_paeth_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x8 = aom_highbd_paeth_predictor_8x8_neon;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_c;
+    if (flags & HAS_NEON) aom_highbd_quantize_b = aom_highbd_quantize_b_neon;
+    aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_neon;
+    aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_c;
+    if (flags & HAS_NEON) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_neon;
+    aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_neon;
+    aom_highbd_quantize_b_64x64_adaptive = aom_highbd_quantize_b_64x64_adaptive_c;
+    if (flags & HAS_NEON) aom_highbd_quantize_b_64x64_adaptive = aom_highbd_quantize_b_64x64_adaptive_neon;
+    aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_c;
+    if (flags & HAS_NEON) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_neon;
+    aom_highbd_sad128x128 = aom_highbd_sad128x128_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x128 = aom_highbd_sad128x128_neon;
+    aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_neon;
+    aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_neon;
+    aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_neon;
+    aom_highbd_sad128x64 = aom_highbd_sad128x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x64 = aom_highbd_sad128x64_neon;
+    aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_neon;
+    aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_neon;
+    aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_neon;
+    aom_highbd_sad16x16 = aom_highbd_sad16x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x16 = aom_highbd_sad16x16_neon;
+    aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_neon;
+    aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_neon;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_neon;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x32 = aom_highbd_sad16x32_neon;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_neon;
+    aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_neon;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_neon;
+    aom_highbd_sad16x4 = aom_highbd_sad16x4_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x4 = aom_highbd_sad16x4_neon;
+    aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_neon;
+    aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_neon;
+    aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_neon;
+    aom_highbd_sad16x64 = aom_highbd_sad16x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x64 = aom_highbd_sad16x64_neon;
+    aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_neon;
+    aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_neon;
+    aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_neon;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x8 = aom_highbd_sad16x8_neon;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_neon;
+    aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_neon;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_neon;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x16 = aom_highbd_sad32x16_neon;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_neon;
+    aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_neon;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_neon;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x32 = aom_highbd_sad32x32_neon;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_neon;
+    aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_neon;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_neon;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x64 = aom_highbd_sad32x64_neon;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_neon;
+    aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_neon;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_neon;
+    aom_highbd_sad32x8 = aom_highbd_sad32x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x8 = aom_highbd_sad32x8_neon;
+    aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_neon;
+    aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_neon;
+    aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_neon;
+    aom_highbd_sad4x16 = aom_highbd_sad4x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x16 = aom_highbd_sad4x16_neon;
+    aom_highbd_sad4x16_avg = aom_highbd_sad4x16_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x16_avg = aom_highbd_sad4x16_avg_neon;
+    aom_highbd_sad4x16x3d = aom_highbd_sad4x16x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x16x3d = aom_highbd_sad4x16x3d_neon;
+    aom_highbd_sad4x16x4d = aom_highbd_sad4x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x16x4d = aom_highbd_sad4x16x4d_neon;
+    aom_highbd_sad4x4 = aom_highbd_sad4x4_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x4 = aom_highbd_sad4x4_neon;
+    aom_highbd_sad4x4_avg = aom_highbd_sad4x4_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x4_avg = aom_highbd_sad4x4_avg_neon;
+    aom_highbd_sad4x4x3d = aom_highbd_sad4x4x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x4x3d = aom_highbd_sad4x4x3d_neon;
+    aom_highbd_sad4x4x4d = aom_highbd_sad4x4x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x4x4d = aom_highbd_sad4x4x4d_neon;
+    aom_highbd_sad4x8 = aom_highbd_sad4x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x8 = aom_highbd_sad4x8_neon;
+    aom_highbd_sad4x8_avg = aom_highbd_sad4x8_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x8_avg = aom_highbd_sad4x8_avg_neon;
+    aom_highbd_sad4x8x3d = aom_highbd_sad4x8x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x8x3d = aom_highbd_sad4x8x3d_neon;
+    aom_highbd_sad4x8x4d = aom_highbd_sad4x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad4x8x4d = aom_highbd_sad4x8x4d_neon;
+    aom_highbd_sad64x128 = aom_highbd_sad64x128_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x128 = aom_highbd_sad64x128_neon;
+    aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_neon;
+    aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_neon;
+    aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_neon;
+    aom_highbd_sad64x16 = aom_highbd_sad64x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x16 = aom_highbd_sad64x16_neon;
+    aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_neon;
+    aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_neon;
+    aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_neon;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x32 = aom_highbd_sad64x32_neon;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_neon;
+    aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_neon;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_neon;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x64 = aom_highbd_sad64x64_neon;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_neon;
+    aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_neon;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_neon;
+    aom_highbd_sad8x16 = aom_highbd_sad8x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x16 = aom_highbd_sad8x16_neon;
+    aom_highbd_sad8x16_avg = aom_highbd_sad8x16_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x16_avg = aom_highbd_sad8x16_avg_neon;
+    aom_highbd_sad8x16x3d = aom_highbd_sad8x16x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x16x3d = aom_highbd_sad8x16x3d_neon;
+    aom_highbd_sad8x16x4d = aom_highbd_sad8x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x16x4d = aom_highbd_sad8x16x4d_neon;
+    aom_highbd_sad8x32 = aom_highbd_sad8x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x32 = aom_highbd_sad8x32_neon;
+    aom_highbd_sad8x32_avg = aom_highbd_sad8x32_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x32_avg = aom_highbd_sad8x32_avg_neon;
+    aom_highbd_sad8x32x3d = aom_highbd_sad8x32x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x32x3d = aom_highbd_sad8x32x3d_neon;
+    aom_highbd_sad8x32x4d = aom_highbd_sad8x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x32x4d = aom_highbd_sad8x32x4d_neon;
+    aom_highbd_sad8x4 = aom_highbd_sad8x4_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x4 = aom_highbd_sad8x4_neon;
+    aom_highbd_sad8x4_avg = aom_highbd_sad8x4_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x4_avg = aom_highbd_sad8x4_avg_neon;
+    aom_highbd_sad8x4x3d = aom_highbd_sad8x4x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x4x3d = aom_highbd_sad8x4x3d_neon;
+    aom_highbd_sad8x4x4d = aom_highbd_sad8x4x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x4x4d = aom_highbd_sad8x4x4d_neon;
+    aom_highbd_sad8x8 = aom_highbd_sad8x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x8 = aom_highbd_sad8x8_neon;
+    aom_highbd_sad8x8_avg = aom_highbd_sad8x8_avg_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x8_avg = aom_highbd_sad8x8_avg_neon;
+    aom_highbd_sad8x8x3d = aom_highbd_sad8x8x3d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x8x3d = aom_highbd_sad8x8x3d_neon;
+    aom_highbd_sad8x8x4d = aom_highbd_sad8x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad8x8x4d = aom_highbd_sad8x8x4d_neon;
+    aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_neon;
+    aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_neon;
+    aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_neon;
+    aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_neon;
+    aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_neon;
+    aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_neon;
+    aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_neon;
+    aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_neon;
+    aom_highbd_sad_skip_16x4 = aom_highbd_sad_skip_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x4 = aom_highbd_sad_skip_16x4_neon;
+    aom_highbd_sad_skip_16x4x4d = aom_highbd_sad_skip_16x4x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x4x4d = aom_highbd_sad_skip_16x4x4d_neon;
+    aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_neon;
+    aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_neon;
+    aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_neon;
+    aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_neon;
+    aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_neon;
+    aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_neon;
+    aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_neon;
+    aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_neon;
+    aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_neon;
+    aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_neon;
+    aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_neon;
+    aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_neon;
+    aom_highbd_sad_skip_4x16 = aom_highbd_sad_skip_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_4x16 = aom_highbd_sad_skip_4x16_neon;
+    aom_highbd_sad_skip_4x16x4d = aom_highbd_sad_skip_4x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_4x16x4d = aom_highbd_sad_skip_4x16x4d_neon;
+    aom_highbd_sad_skip_4x4 = aom_highbd_sad_skip_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_4x4 = aom_highbd_sad_skip_4x4_neon;
+    aom_highbd_sad_skip_4x4x4d = aom_highbd_sad_skip_4x4x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_4x4x4d = aom_highbd_sad_skip_4x4x4d_neon;
+    aom_highbd_sad_skip_4x8 = aom_highbd_sad_skip_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_4x8 = aom_highbd_sad_skip_4x8_neon;
+    aom_highbd_sad_skip_4x8x4d = aom_highbd_sad_skip_4x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_4x8x4d = aom_highbd_sad_skip_4x8x4d_neon;
+    aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_neon;
+    aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_neon;
+    aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_neon;
+    aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_neon;
+    aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_neon;
+    aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_neon;
+    aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_neon;
+    aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_neon;
+    aom_highbd_sad_skip_8x16 = aom_highbd_sad_skip_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x16 = aom_highbd_sad_skip_8x16_neon;
+    aom_highbd_sad_skip_8x16x4d = aom_highbd_sad_skip_8x16x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x16x4d = aom_highbd_sad_skip_8x16x4d_neon;
+    aom_highbd_sad_skip_8x32 = aom_highbd_sad_skip_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x32 = aom_highbd_sad_skip_8x32_neon;
+    aom_highbd_sad_skip_8x32x4d = aom_highbd_sad_skip_8x32x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x32x4d = aom_highbd_sad_skip_8x32x4d_neon;
+    aom_highbd_sad_skip_8x4 = aom_highbd_sad_skip_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x4 = aom_highbd_sad_skip_8x4_neon;
+    aom_highbd_sad_skip_8x4x4d = aom_highbd_sad_skip_8x4x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x4x4d = aom_highbd_sad_skip_8x4x4d_neon;
+    aom_highbd_sad_skip_8x8 = aom_highbd_sad_skip_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x8 = aom_highbd_sad_skip_8x8_neon;
+    aom_highbd_sad_skip_8x8x4d = aom_highbd_sad_skip_8x8x4d_c;
+    if (flags & HAS_NEON) aom_highbd_sad_skip_8x8x4d = aom_highbd_sad_skip_8x8x4d_neon;
+    aom_highbd_smooth_h_predictor_16x16 = aom_highbd_smooth_h_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x16 = aom_highbd_smooth_h_predictor_16x16_neon;
+    aom_highbd_smooth_h_predictor_16x32 = aom_highbd_smooth_h_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x32 = aom_highbd_smooth_h_predictor_16x32_neon;
+    aom_highbd_smooth_h_predictor_16x4 = aom_highbd_smooth_h_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x4 = aom_highbd_smooth_h_predictor_16x4_neon;
+    aom_highbd_smooth_h_predictor_16x64 = aom_highbd_smooth_h_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x64 = aom_highbd_smooth_h_predictor_16x64_neon;
+    aom_highbd_smooth_h_predictor_16x8 = aom_highbd_smooth_h_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x8 = aom_highbd_smooth_h_predictor_16x8_neon;
+    aom_highbd_smooth_h_predictor_32x16 = aom_highbd_smooth_h_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x16 = aom_highbd_smooth_h_predictor_32x16_neon;
+    aom_highbd_smooth_h_predictor_32x32 = aom_highbd_smooth_h_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x32 = aom_highbd_smooth_h_predictor_32x32_neon;
+    aom_highbd_smooth_h_predictor_32x64 = aom_highbd_smooth_h_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x64 = aom_highbd_smooth_h_predictor_32x64_neon;
+    aom_highbd_smooth_h_predictor_32x8 = aom_highbd_smooth_h_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x8 = aom_highbd_smooth_h_predictor_32x8_neon;
+    aom_highbd_smooth_h_predictor_4x16 = aom_highbd_smooth_h_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_4x16 = aom_highbd_smooth_h_predictor_4x16_neon;
+    aom_highbd_smooth_h_predictor_4x4 = aom_highbd_smooth_h_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_4x4 = aom_highbd_smooth_h_predictor_4x4_neon;
+    aom_highbd_smooth_h_predictor_4x8 = aom_highbd_smooth_h_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_4x8 = aom_highbd_smooth_h_predictor_4x8_neon;
+    aom_highbd_smooth_h_predictor_64x16 = aom_highbd_smooth_h_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_64x16 = aom_highbd_smooth_h_predictor_64x16_neon;
+    aom_highbd_smooth_h_predictor_64x32 = aom_highbd_smooth_h_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_64x32 = aom_highbd_smooth_h_predictor_64x32_neon;
+    aom_highbd_smooth_h_predictor_64x64 = aom_highbd_smooth_h_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_64x64 = aom_highbd_smooth_h_predictor_64x64_neon;
+    aom_highbd_smooth_h_predictor_8x16 = aom_highbd_smooth_h_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x16 = aom_highbd_smooth_h_predictor_8x16_neon;
+    aom_highbd_smooth_h_predictor_8x32 = aom_highbd_smooth_h_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x32 = aom_highbd_smooth_h_predictor_8x32_neon;
+    aom_highbd_smooth_h_predictor_8x4 = aom_highbd_smooth_h_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x4 = aom_highbd_smooth_h_predictor_8x4_neon;
+    aom_highbd_smooth_h_predictor_8x8 = aom_highbd_smooth_h_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x8 = aom_highbd_smooth_h_predictor_8x8_neon;
+    aom_highbd_smooth_predictor_16x16 = aom_highbd_smooth_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x16 = aom_highbd_smooth_predictor_16x16_neon;
+    aom_highbd_smooth_predictor_16x32 = aom_highbd_smooth_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x32 = aom_highbd_smooth_predictor_16x32_neon;
+    aom_highbd_smooth_predictor_16x4 = aom_highbd_smooth_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x4 = aom_highbd_smooth_predictor_16x4_neon;
+    aom_highbd_smooth_predictor_16x64 = aom_highbd_smooth_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x64 = aom_highbd_smooth_predictor_16x64_neon;
+    aom_highbd_smooth_predictor_16x8 = aom_highbd_smooth_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x8 = aom_highbd_smooth_predictor_16x8_neon;
+    aom_highbd_smooth_predictor_32x16 = aom_highbd_smooth_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x16 = aom_highbd_smooth_predictor_32x16_neon;
+    aom_highbd_smooth_predictor_32x32 = aom_highbd_smooth_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x32 = aom_highbd_smooth_predictor_32x32_neon;
+    aom_highbd_smooth_predictor_32x64 = aom_highbd_smooth_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x64 = aom_highbd_smooth_predictor_32x64_neon;
+    aom_highbd_smooth_predictor_32x8 = aom_highbd_smooth_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x8 = aom_highbd_smooth_predictor_32x8_neon;
+    aom_highbd_smooth_predictor_4x16 = aom_highbd_smooth_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_4x16 = aom_highbd_smooth_predictor_4x16_neon;
+    aom_highbd_smooth_predictor_4x4 = aom_highbd_smooth_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_4x4 = aom_highbd_smooth_predictor_4x4_neon;
+    aom_highbd_smooth_predictor_4x8 = aom_highbd_smooth_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_4x8 = aom_highbd_smooth_predictor_4x8_neon;
+    aom_highbd_smooth_predictor_64x16 = aom_highbd_smooth_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_64x16 = aom_highbd_smooth_predictor_64x16_neon;
+    aom_highbd_smooth_predictor_64x32 = aom_highbd_smooth_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_64x32 = aom_highbd_smooth_predictor_64x32_neon;
+    aom_highbd_smooth_predictor_64x64 = aom_highbd_smooth_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_64x64 = aom_highbd_smooth_predictor_64x64_neon;
+    aom_highbd_smooth_predictor_8x16 = aom_highbd_smooth_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x16 = aom_highbd_smooth_predictor_8x16_neon;
+    aom_highbd_smooth_predictor_8x32 = aom_highbd_smooth_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x32 = aom_highbd_smooth_predictor_8x32_neon;
+    aom_highbd_smooth_predictor_8x4 = aom_highbd_smooth_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x4 = aom_highbd_smooth_predictor_8x4_neon;
+    aom_highbd_smooth_predictor_8x8 = aom_highbd_smooth_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x8 = aom_highbd_smooth_predictor_8x8_neon;
+    aom_highbd_smooth_v_predictor_16x16 = aom_highbd_smooth_v_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x16 = aom_highbd_smooth_v_predictor_16x16_neon;
+    aom_highbd_smooth_v_predictor_16x32 = aom_highbd_smooth_v_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x32 = aom_highbd_smooth_v_predictor_16x32_neon;
+    aom_highbd_smooth_v_predictor_16x4 = aom_highbd_smooth_v_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x4 = aom_highbd_smooth_v_predictor_16x4_neon;
+    aom_highbd_smooth_v_predictor_16x64 = aom_highbd_smooth_v_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x64 = aom_highbd_smooth_v_predictor_16x64_neon;
+    aom_highbd_smooth_v_predictor_16x8 = aom_highbd_smooth_v_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x8 = aom_highbd_smooth_v_predictor_16x8_neon;
+    aom_highbd_smooth_v_predictor_32x16 = aom_highbd_smooth_v_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x16 = aom_highbd_smooth_v_predictor_32x16_neon;
+    aom_highbd_smooth_v_predictor_32x32 = aom_highbd_smooth_v_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x32 = aom_highbd_smooth_v_predictor_32x32_neon;
+    aom_highbd_smooth_v_predictor_32x64 = aom_highbd_smooth_v_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x64 = aom_highbd_smooth_v_predictor_32x64_neon;
+    aom_highbd_smooth_v_predictor_32x8 = aom_highbd_smooth_v_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x8 = aom_highbd_smooth_v_predictor_32x8_neon;
+    aom_highbd_smooth_v_predictor_4x16 = aom_highbd_smooth_v_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_4x16 = aom_highbd_smooth_v_predictor_4x16_neon;
+    aom_highbd_smooth_v_predictor_4x4 = aom_highbd_smooth_v_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_4x4 = aom_highbd_smooth_v_predictor_4x4_neon;
+    aom_highbd_smooth_v_predictor_4x8 = aom_highbd_smooth_v_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_4x8 = aom_highbd_smooth_v_predictor_4x8_neon;
+    aom_highbd_smooth_v_predictor_64x16 = aom_highbd_smooth_v_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_64x16 = aom_highbd_smooth_v_predictor_64x16_neon;
+    aom_highbd_smooth_v_predictor_64x32 = aom_highbd_smooth_v_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_64x32 = aom_highbd_smooth_v_predictor_64x32_neon;
+    aom_highbd_smooth_v_predictor_64x64 = aom_highbd_smooth_v_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_64x64 = aom_highbd_smooth_v_predictor_64x64_neon;
+    aom_highbd_smooth_v_predictor_8x16 = aom_highbd_smooth_v_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x16 = aom_highbd_smooth_v_predictor_8x16_neon;
+    aom_highbd_smooth_v_predictor_8x32 = aom_highbd_smooth_v_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x32 = aom_highbd_smooth_v_predictor_8x32_neon;
+    aom_highbd_smooth_v_predictor_8x4 = aom_highbd_smooth_v_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x4 = aom_highbd_smooth_v_predictor_8x4_neon;
+    aom_highbd_smooth_v_predictor_8x8 = aom_highbd_smooth_v_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x8 = aom_highbd_smooth_v_predictor_8x8_neon;
+    aom_highbd_sse = aom_highbd_sse_c;
+    if (flags & HAS_NEON) aom_highbd_sse = aom_highbd_sse_neon;
+    aom_highbd_subtract_block = aom_highbd_subtract_block_c;
+    if (flags & HAS_NEON) aom_highbd_subtract_block = aom_highbd_subtract_block_neon;
+    aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_neon;
+    aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_neon;
+    aom_highbd_v_predictor_16x4 = aom_highbd_v_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_16x4 = aom_highbd_v_predictor_16x4_neon;
+    aom_highbd_v_predictor_16x64 = aom_highbd_v_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_16x64 = aom_highbd_v_predictor_16x64_neon;
+    aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_neon;
+    aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_neon;
+    aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_neon;
+    aom_highbd_v_predictor_32x64 = aom_highbd_v_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_32x64 = aom_highbd_v_predictor_32x64_neon;
+    aom_highbd_v_predictor_32x8 = aom_highbd_v_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_32x8 = aom_highbd_v_predictor_32x8_neon;
+    aom_highbd_v_predictor_4x16 = aom_highbd_v_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_4x16 = aom_highbd_v_predictor_4x16_neon;
+    aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_neon;
+    aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_neon;
+    aom_highbd_v_predictor_64x16 = aom_highbd_v_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_64x16 = aom_highbd_v_predictor_64x16_neon;
+    aom_highbd_v_predictor_64x32 = aom_highbd_v_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_64x32 = aom_highbd_v_predictor_64x32_neon;
+    aom_highbd_v_predictor_64x64 = aom_highbd_v_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_64x64 = aom_highbd_v_predictor_64x64_neon;
+    aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_neon;
+    aom_highbd_v_predictor_8x32 = aom_highbd_v_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_8x32 = aom_highbd_v_predictor_8x32_neon;
+    aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_neon;
+    aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_neon;
+    aom_int_pro_col = aom_int_pro_col_c;
+    if (flags & HAS_NEON) aom_int_pro_col = aom_int_pro_col_neon;
+    aom_int_pro_row = aom_int_pro_row_c;
+    if (flags & HAS_NEON) aom_int_pro_row = aom_int_pro_row_neon;
+    aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
+    if (flags & HAS_NEON) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_neon;
+    aom_lpf_horizontal_14 = aom_lpf_horizontal_14_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_neon;
+    aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_neon;
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_neon;
+    aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_neon;
+    aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_neon;
+    aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_neon;
+    aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_neon;
+    aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_neon;
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_neon;
+    aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_neon;
+    aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_neon;
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_neon;
+    aom_lpf_vertical_14 = aom_lpf_vertical_14_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_14 = aom_lpf_vertical_14_neon;
+    aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_neon;
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_neon;
+    aom_lpf_vertical_4 = aom_lpf_vertical_4_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_4 = aom_lpf_vertical_4_neon;
+    aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_neon;
+    aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_neon;
+    aom_lpf_vertical_6 = aom_lpf_vertical_6_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_6 = aom_lpf_vertical_6_neon;
+    aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_neon;
+    aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_neon;
+    aom_lpf_vertical_8 = aom_lpf_vertical_8_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_8 = aom_lpf_vertical_8_neon;
+    aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_neon;
+    aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_neon;
+    aom_masked_sad128x128 = aom_masked_sad128x128_c;
+    if (flags & HAS_NEON) aom_masked_sad128x128 = aom_masked_sad128x128_neon;
+    aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_neon;
+    aom_masked_sad128x64 = aom_masked_sad128x64_c;
+    if (flags & HAS_NEON) aom_masked_sad128x64 = aom_masked_sad128x64_neon;
+    aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_neon;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_NEON) aom_masked_sad16x16 = aom_masked_sad16x16_neon;
+    aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_neon;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_NEON) aom_masked_sad16x32 = aom_masked_sad16x32_neon;
+    aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_neon;
+    aom_masked_sad16x4 = aom_masked_sad16x4_c;
+    if (flags & HAS_NEON) aom_masked_sad16x4 = aom_masked_sad16x4_neon;
+    aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_neon;
+    aom_masked_sad16x64 = aom_masked_sad16x64_c;
+    if (flags & HAS_NEON) aom_masked_sad16x64 = aom_masked_sad16x64_neon;
+    aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_neon;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_NEON) aom_masked_sad16x8 = aom_masked_sad16x8_neon;
+    aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_neon;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_NEON) aom_masked_sad32x16 = aom_masked_sad32x16_neon;
+    aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_neon;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_NEON) aom_masked_sad32x32 = aom_masked_sad32x32_neon;
+    aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_neon;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_NEON) aom_masked_sad32x64 = aom_masked_sad32x64_neon;
+    aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_neon;
+    aom_masked_sad32x8 = aom_masked_sad32x8_c;
+    if (flags & HAS_NEON) aom_masked_sad32x8 = aom_masked_sad32x8_neon;
+    aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_neon;
+    aom_masked_sad4x16 = aom_masked_sad4x16_c;
+    if (flags & HAS_NEON) aom_masked_sad4x16 = aom_masked_sad4x16_neon;
+    aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_neon;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_NEON) aom_masked_sad4x4 = aom_masked_sad4x4_neon;
+    aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_neon;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_NEON) aom_masked_sad4x8 = aom_masked_sad4x8_neon;
+    aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_neon;
+    aom_masked_sad64x128 = aom_masked_sad64x128_c;
+    if (flags & HAS_NEON) aom_masked_sad64x128 = aom_masked_sad64x128_neon;
+    aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_neon;
+    aom_masked_sad64x16 = aom_masked_sad64x16_c;
+    if (flags & HAS_NEON) aom_masked_sad64x16 = aom_masked_sad64x16_neon;
+    aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_neon;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_NEON) aom_masked_sad64x32 = aom_masked_sad64x32_neon;
+    aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_neon;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_NEON) aom_masked_sad64x64 = aom_masked_sad64x64_neon;
+    aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_neon;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_NEON) aom_masked_sad8x16 = aom_masked_sad8x16_neon;
+    aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_neon;
+    aom_masked_sad8x32 = aom_masked_sad8x32_c;
+    if (flags & HAS_NEON) aom_masked_sad8x32 = aom_masked_sad8x32_neon;
+    aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_neon;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_NEON) aom_masked_sad8x4 = aom_masked_sad8x4_neon;
+    aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_neon;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_NEON) aom_masked_sad8x8 = aom_masked_sad8x8_neon;
+    aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_c;
+    if (flags & HAS_NEON) aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_neon;
+    aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_neon;
+    aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_neon;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_neon;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_neon;
+    aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_neon;
+    aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_neon;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_neon;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_neon;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_neon;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_neon;
+    aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_neon;
+    aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_neon;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_neon;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_neon;
+    aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_neon;
+    aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_neon;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_neon;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_neon;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_neon;
+    aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_neon;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_neon;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_neon;
+    aom_minmax_8x8 = aom_minmax_8x8_c;
+    if (flags & HAS_NEON) aom_minmax_8x8 = aom_minmax_8x8_neon;
+    aom_mse16x16 = aom_mse16x16_c;
+    if (flags & HAS_NEON) aom_mse16x16 = aom_mse16x16_neon;
+    aom_mse16x8 = aom_mse16x8_c;
+    if (flags & HAS_NEON) aom_mse16x8 = aom_mse16x8_neon;
+    aom_mse8x16 = aom_mse8x16_c;
+    if (flags & HAS_NEON) aom_mse8x16 = aom_mse8x16_neon;
+    aom_mse8x8 = aom_mse8x8_c;
+    if (flags & HAS_NEON) aom_mse8x8 = aom_mse8x8_neon;
+    aom_mse_16xh_16bit = aom_mse_16xh_16bit_c;
+    if (flags & HAS_NEON) aom_mse_16xh_16bit = aom_mse_16xh_16bit_neon;
+    aom_mse_wxh_16bit = aom_mse_wxh_16bit_c;
+    if (flags & HAS_NEON) aom_mse_wxh_16bit = aom_mse_wxh_16bit_neon;
+    aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_c;
+    if (flags & HAS_NEON) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_neon;
+    aom_obmc_sad128x128 = aom_obmc_sad128x128_c;
+    if (flags & HAS_NEON) aom_obmc_sad128x128 = aom_obmc_sad128x128_neon;
+    aom_obmc_sad128x64 = aom_obmc_sad128x64_c;
+    if (flags & HAS_NEON) aom_obmc_sad128x64 = aom_obmc_sad128x64_neon;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_NEON) aom_obmc_sad16x16 = aom_obmc_sad16x16_neon;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_NEON) aom_obmc_sad16x32 = aom_obmc_sad16x32_neon;
+    aom_obmc_sad16x4 = aom_obmc_sad16x4_c;
+    if (flags & HAS_NEON) aom_obmc_sad16x4 = aom_obmc_sad16x4_neon;
+    aom_obmc_sad16x64 = aom_obmc_sad16x64_c;
+    if (flags & HAS_NEON) aom_obmc_sad16x64 = aom_obmc_sad16x64_neon;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_NEON) aom_obmc_sad16x8 = aom_obmc_sad16x8_neon;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_NEON) aom_obmc_sad32x16 = aom_obmc_sad32x16_neon;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_NEON) aom_obmc_sad32x32 = aom_obmc_sad32x32_neon;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_NEON) aom_obmc_sad32x64 = aom_obmc_sad32x64_neon;
+    aom_obmc_sad32x8 = aom_obmc_sad32x8_c;
+    if (flags & HAS_NEON) aom_obmc_sad32x8 = aom_obmc_sad32x8_neon;
+    aom_obmc_sad4x16 = aom_obmc_sad4x16_c;
+    if (flags & HAS_NEON) aom_obmc_sad4x16 = aom_obmc_sad4x16_neon;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_NEON) aom_obmc_sad4x4 = aom_obmc_sad4x4_neon;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_NEON) aom_obmc_sad4x8 = aom_obmc_sad4x8_neon;
+    aom_obmc_sad64x128 = aom_obmc_sad64x128_c;
+    if (flags & HAS_NEON) aom_obmc_sad64x128 = aom_obmc_sad64x128_neon;
+    aom_obmc_sad64x16 = aom_obmc_sad64x16_c;
+    if (flags & HAS_NEON) aom_obmc_sad64x16 = aom_obmc_sad64x16_neon;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_NEON) aom_obmc_sad64x32 = aom_obmc_sad64x32_neon;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_NEON) aom_obmc_sad64x64 = aom_obmc_sad64x64_neon;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_NEON) aom_obmc_sad8x16 = aom_obmc_sad8x16_neon;
+    aom_obmc_sad8x32 = aom_obmc_sad8x32_c;
+    if (flags & HAS_NEON) aom_obmc_sad8x32 = aom_obmc_sad8x32_neon;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_NEON) aom_obmc_sad8x4 = aom_obmc_sad8x4_neon;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_NEON) aom_obmc_sad8x8 = aom_obmc_sad8x8_neon;
+    aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_neon;
+    aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_neon;
+    aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_neon;
+    aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_neon;
+    aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_neon;
+    aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_neon;
+    aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_neon;
+    aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_neon;
+    aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_neon;
+    aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_neon;
+    aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_neon;
+    aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_neon;
+    aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_neon;
+    aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_neon;
+    aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_neon;
+    aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_neon;
+    aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_neon;
+    aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_neon;
+    aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_neon;
+    aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_neon;
+    aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_neon;
+    aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_neon;
+    aom_obmc_variance128x128 = aom_obmc_variance128x128_c;
+    if (flags & HAS_NEON) aom_obmc_variance128x128 = aom_obmc_variance128x128_neon;
+    aom_obmc_variance128x64 = aom_obmc_variance128x64_c;
+    if (flags & HAS_NEON) aom_obmc_variance128x64 = aom_obmc_variance128x64_neon;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_NEON) aom_obmc_variance16x16 = aom_obmc_variance16x16_neon;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_NEON) aom_obmc_variance16x32 = aom_obmc_variance16x32_neon;
+    aom_obmc_variance16x4 = aom_obmc_variance16x4_c;
+    if (flags & HAS_NEON) aom_obmc_variance16x4 = aom_obmc_variance16x4_neon;
+    aom_obmc_variance16x64 = aom_obmc_variance16x64_c;
+    if (flags & HAS_NEON) aom_obmc_variance16x64 = aom_obmc_variance16x64_neon;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_NEON) aom_obmc_variance16x8 = aom_obmc_variance16x8_neon;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_NEON) aom_obmc_variance32x16 = aom_obmc_variance32x16_neon;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_NEON) aom_obmc_variance32x32 = aom_obmc_variance32x32_neon;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_NEON) aom_obmc_variance32x64 = aom_obmc_variance32x64_neon;
+    aom_obmc_variance32x8 = aom_obmc_variance32x8_c;
+    if (flags & HAS_NEON) aom_obmc_variance32x8 = aom_obmc_variance32x8_neon;
+    aom_obmc_variance4x16 = aom_obmc_variance4x16_c;
+    if (flags & HAS_NEON) aom_obmc_variance4x16 = aom_obmc_variance4x16_neon;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_NEON) aom_obmc_variance4x4 = aom_obmc_variance4x4_neon;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_NEON) aom_obmc_variance4x8 = aom_obmc_variance4x8_neon;
+    aom_obmc_variance64x128 = aom_obmc_variance64x128_c;
+    if (flags & HAS_NEON) aom_obmc_variance64x128 = aom_obmc_variance64x128_neon;
+    aom_obmc_variance64x16 = aom_obmc_variance64x16_c;
+    if (flags & HAS_NEON) aom_obmc_variance64x16 = aom_obmc_variance64x16_neon;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_NEON) aom_obmc_variance64x32 = aom_obmc_variance64x32_neon;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_NEON) aom_obmc_variance64x64 = aom_obmc_variance64x64_neon;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_NEON) aom_obmc_variance8x16 = aom_obmc_variance8x16_neon;
+    aom_obmc_variance8x32 = aom_obmc_variance8x32_c;
+    if (flags & HAS_NEON) aom_obmc_variance8x32 = aom_obmc_variance8x32_neon;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_NEON) aom_obmc_variance8x4 = aom_obmc_variance8x4_neon;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_NEON) aom_obmc_variance8x8 = aom_obmc_variance8x8_neon;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_neon;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_neon;
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_neon;
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_neon;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_neon;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_neon;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_neon;
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_neon;
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_neon;
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_neon;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_neon;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_neon;
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_neon;
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_neon;
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_neon;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_neon;
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_neon;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_neon;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_neon;
+    aom_quantize_b = aom_quantize_b_c;
+    if (flags & HAS_NEON) aom_quantize_b = aom_quantize_b_neon;
+    aom_quantize_b_32x32 = aom_quantize_b_32x32_c;
+    if (flags & HAS_NEON) aom_quantize_b_32x32 = aom_quantize_b_32x32_neon;
+    aom_quantize_b_64x64 = aom_quantize_b_64x64_c;
+    if (flags & HAS_NEON) aom_quantize_b_64x64 = aom_quantize_b_64x64_neon;
+    aom_sad128x128 = aom_sad128x128_c;
+    if (flags & HAS_NEON) aom_sad128x128 = aom_sad128x128_neon;
+    aom_sad128x128_avg = aom_sad128x128_avg_c;
+    if (flags & HAS_NEON) aom_sad128x128_avg = aom_sad128x128_avg_neon;
+    aom_sad128x128x3d = aom_sad128x128x3d_c;
+    if (flags & HAS_NEON) aom_sad128x128x3d = aom_sad128x128x3d_neon;
+    aom_sad128x128x4d = aom_sad128x128x4d_c;
+    if (flags & HAS_NEON) aom_sad128x128x4d = aom_sad128x128x4d_neon;
+    aom_sad128x64 = aom_sad128x64_c;
+    if (flags & HAS_NEON) aom_sad128x64 = aom_sad128x64_neon;
+    aom_sad128x64_avg = aom_sad128x64_avg_c;
+    if (flags & HAS_NEON) aom_sad128x64_avg = aom_sad128x64_avg_neon;
+    aom_sad128x64x3d = aom_sad128x64x3d_c;
+    if (flags & HAS_NEON) aom_sad128x64x3d = aom_sad128x64x3d_neon;
+    aom_sad128x64x4d = aom_sad128x64x4d_c;
+    if (flags & HAS_NEON) aom_sad128x64x4d = aom_sad128x64x4d_neon;
+    aom_sad16x16 = aom_sad16x16_c;
+    if (flags & HAS_NEON) aom_sad16x16 = aom_sad16x16_neon;
+    aom_sad16x16_avg = aom_sad16x16_avg_c;
+    if (flags & HAS_NEON) aom_sad16x16_avg = aom_sad16x16_avg_neon;
+    aom_sad16x16x3d = aom_sad16x16x3d_c;
+    if (flags & HAS_NEON) aom_sad16x16x3d = aom_sad16x16x3d_neon;
+    aom_sad16x16x4d = aom_sad16x16x4d_c;
+    if (flags & HAS_NEON) aom_sad16x16x4d = aom_sad16x16x4d_neon;
+    aom_sad16x32 = aom_sad16x32_c;
+    if (flags & HAS_NEON) aom_sad16x32 = aom_sad16x32_neon;
+    aom_sad16x32_avg = aom_sad16x32_avg_c;
+    if (flags & HAS_NEON) aom_sad16x32_avg = aom_sad16x32_avg_neon;
+    aom_sad16x32x3d = aom_sad16x32x3d_c;
+    if (flags & HAS_NEON) aom_sad16x32x3d = aom_sad16x32x3d_neon;
+    aom_sad16x32x4d = aom_sad16x32x4d_c;
+    if (flags & HAS_NEON) aom_sad16x32x4d = aom_sad16x32x4d_neon;
+    aom_sad16x4 = aom_sad16x4_c;
+    if (flags & HAS_NEON) aom_sad16x4 = aom_sad16x4_neon;
+    aom_sad16x4_avg = aom_sad16x4_avg_c;
+    if (flags & HAS_NEON) aom_sad16x4_avg = aom_sad16x4_avg_neon;
+    aom_sad16x4x3d = aom_sad16x4x3d_c;
+    if (flags & HAS_NEON) aom_sad16x4x3d = aom_sad16x4x3d_neon;
+    aom_sad16x4x4d = aom_sad16x4x4d_c;
+    if (flags & HAS_NEON) aom_sad16x4x4d = aom_sad16x4x4d_neon;
+    aom_sad16x64 = aom_sad16x64_c;
+    if (flags & HAS_NEON) aom_sad16x64 = aom_sad16x64_neon;
+    aom_sad16x64_avg = aom_sad16x64_avg_c;
+    if (flags & HAS_NEON) aom_sad16x64_avg = aom_sad16x64_avg_neon;
+    aom_sad16x64x3d = aom_sad16x64x3d_c;
+    if (flags & HAS_NEON) aom_sad16x64x3d = aom_sad16x64x3d_neon;
+    aom_sad16x64x4d = aom_sad16x64x4d_c;
+    if (flags & HAS_NEON) aom_sad16x64x4d = aom_sad16x64x4d_neon;
+    aom_sad16x8 = aom_sad16x8_c;
+    if (flags & HAS_NEON) aom_sad16x8 = aom_sad16x8_neon;
+    aom_sad16x8_avg = aom_sad16x8_avg_c;
+    if (flags & HAS_NEON) aom_sad16x8_avg = aom_sad16x8_avg_neon;
+    aom_sad16x8x3d = aom_sad16x8x3d_c;
+    if (flags & HAS_NEON) aom_sad16x8x3d = aom_sad16x8x3d_neon;
+    aom_sad16x8x4d = aom_sad16x8x4d_c;
+    if (flags & HAS_NEON) aom_sad16x8x4d = aom_sad16x8x4d_neon;
+    aom_sad32x16 = aom_sad32x16_c;
+    if (flags & HAS_NEON) aom_sad32x16 = aom_sad32x16_neon;
+    aom_sad32x16_avg = aom_sad32x16_avg_c;
+    if (flags & HAS_NEON) aom_sad32x16_avg = aom_sad32x16_avg_neon;
+    aom_sad32x16x3d = aom_sad32x16x3d_c;
+    if (flags & HAS_NEON) aom_sad32x16x3d = aom_sad32x16x3d_neon;
+    aom_sad32x16x4d = aom_sad32x16x4d_c;
+    if (flags & HAS_NEON) aom_sad32x16x4d = aom_sad32x16x4d_neon;
+    aom_sad32x32 = aom_sad32x32_c;
+    if (flags & HAS_NEON) aom_sad32x32 = aom_sad32x32_neon;
+    aom_sad32x32_avg = aom_sad32x32_avg_c;
+    if (flags & HAS_NEON) aom_sad32x32_avg = aom_sad32x32_avg_neon;
+    aom_sad32x32x3d = aom_sad32x32x3d_c;
+    if (flags & HAS_NEON) aom_sad32x32x3d = aom_sad32x32x3d_neon;
+    aom_sad32x32x4d = aom_sad32x32x4d_c;
+    if (flags & HAS_NEON) aom_sad32x32x4d = aom_sad32x32x4d_neon;
+    aom_sad32x64 = aom_sad32x64_c;
+    if (flags & HAS_NEON) aom_sad32x64 = aom_sad32x64_neon;
+    aom_sad32x64_avg = aom_sad32x64_avg_c;
+    if (flags & HAS_NEON) aom_sad32x64_avg = aom_sad32x64_avg_neon;
+    aom_sad32x64x3d = aom_sad32x64x3d_c;
+    if (flags & HAS_NEON) aom_sad32x64x3d = aom_sad32x64x3d_neon;
+    aom_sad32x64x4d = aom_sad32x64x4d_c;
+    if (flags & HAS_NEON) aom_sad32x64x4d = aom_sad32x64x4d_neon;
+    aom_sad32x8 = aom_sad32x8_c;
+    if (flags & HAS_NEON) aom_sad32x8 = aom_sad32x8_neon;
+    aom_sad32x8_avg = aom_sad32x8_avg_c;
+    if (flags & HAS_NEON) aom_sad32x8_avg = aom_sad32x8_avg_neon;
+    aom_sad32x8x3d = aom_sad32x8x3d_c;
+    if (flags & HAS_NEON) aom_sad32x8x3d = aom_sad32x8x3d_neon;
+    aom_sad32x8x4d = aom_sad32x8x4d_c;
+    if (flags & HAS_NEON) aom_sad32x8x4d = aom_sad32x8x4d_neon;
+    aom_sad4x16 = aom_sad4x16_c;
+    if (flags & HAS_NEON) aom_sad4x16 = aom_sad4x16_neon;
+    aom_sad4x16_avg = aom_sad4x16_avg_c;
+    if (flags & HAS_NEON) aom_sad4x16_avg = aom_sad4x16_avg_neon;
+    aom_sad4x16x3d = aom_sad4x16x3d_c;
+    if (flags & HAS_NEON) aom_sad4x16x3d = aom_sad4x16x3d_neon;
+    aom_sad4x16x4d = aom_sad4x16x4d_c;
+    if (flags & HAS_NEON) aom_sad4x16x4d = aom_sad4x16x4d_neon;
+    aom_sad4x4 = aom_sad4x4_c;
+    if (flags & HAS_NEON) aom_sad4x4 = aom_sad4x4_neon;
+    aom_sad4x4_avg = aom_sad4x4_avg_c;
+    if (flags & HAS_NEON) aom_sad4x4_avg = aom_sad4x4_avg_neon;
+    aom_sad4x4x3d = aom_sad4x4x3d_c;
+    if (flags & HAS_NEON) aom_sad4x4x3d = aom_sad4x4x3d_neon;
+    aom_sad4x4x4d = aom_sad4x4x4d_c;
+    if (flags & HAS_NEON) aom_sad4x4x4d = aom_sad4x4x4d_neon;
+    aom_sad4x8 = aom_sad4x8_c;
+    if (flags & HAS_NEON) aom_sad4x8 = aom_sad4x8_neon;
+    aom_sad4x8_avg = aom_sad4x8_avg_c;
+    if (flags & HAS_NEON) aom_sad4x8_avg = aom_sad4x8_avg_neon;
+    aom_sad4x8x3d = aom_sad4x8x3d_c;
+    if (flags & HAS_NEON) aom_sad4x8x3d = aom_sad4x8x3d_neon;
+    aom_sad4x8x4d = aom_sad4x8x4d_c;
+    if (flags & HAS_NEON) aom_sad4x8x4d = aom_sad4x8x4d_neon;
+    aom_sad64x128 = aom_sad64x128_c;
+    if (flags & HAS_NEON) aom_sad64x128 = aom_sad64x128_neon;
+    aom_sad64x128_avg = aom_sad64x128_avg_c;
+    if (flags & HAS_NEON) aom_sad64x128_avg = aom_sad64x128_avg_neon;
+    aom_sad64x128x3d = aom_sad64x128x3d_c;
+    if (flags & HAS_NEON) aom_sad64x128x3d = aom_sad64x128x3d_neon;
+    aom_sad64x128x4d = aom_sad64x128x4d_c;
+    if (flags & HAS_NEON) aom_sad64x128x4d = aom_sad64x128x4d_neon;
+    aom_sad64x16 = aom_sad64x16_c;
+    if (flags & HAS_NEON) aom_sad64x16 = aom_sad64x16_neon;
+    aom_sad64x16_avg = aom_sad64x16_avg_c;
+    if (flags & HAS_NEON) aom_sad64x16_avg = aom_sad64x16_avg_neon;
+    aom_sad64x16x3d = aom_sad64x16x3d_c;
+    if (flags & HAS_NEON) aom_sad64x16x3d = aom_sad64x16x3d_neon;
+    aom_sad64x16x4d = aom_sad64x16x4d_c;
+    if (flags & HAS_NEON) aom_sad64x16x4d = aom_sad64x16x4d_neon;
+    aom_sad64x32 = aom_sad64x32_c;
+    if (flags & HAS_NEON) aom_sad64x32 = aom_sad64x32_neon;
+    aom_sad64x32_avg = aom_sad64x32_avg_c;
+    if (flags & HAS_NEON) aom_sad64x32_avg = aom_sad64x32_avg_neon;
+    aom_sad64x32x3d = aom_sad64x32x3d_c;
+    if (flags & HAS_NEON) aom_sad64x32x3d = aom_sad64x32x3d_neon;
+    aom_sad64x32x4d = aom_sad64x32x4d_c;
+    if (flags & HAS_NEON) aom_sad64x32x4d = aom_sad64x32x4d_neon;
+    aom_sad64x64 = aom_sad64x64_c;
+    if (flags & HAS_NEON) aom_sad64x64 = aom_sad64x64_neon;
+    aom_sad64x64_avg = aom_sad64x64_avg_c;
+    if (flags & HAS_NEON) aom_sad64x64_avg = aom_sad64x64_avg_neon;
+    aom_sad64x64x3d = aom_sad64x64x3d_c;
+    if (flags & HAS_NEON) aom_sad64x64x3d = aom_sad64x64x3d_neon;
+    aom_sad64x64x4d = aom_sad64x64x4d_c;
+    if (flags & HAS_NEON) aom_sad64x64x4d = aom_sad64x64x4d_neon;
+    aom_sad8x16 = aom_sad8x16_c;
+    if (flags & HAS_NEON) aom_sad8x16 = aom_sad8x16_neon;
+    aom_sad8x16_avg = aom_sad8x16_avg_c;
+    if (flags & HAS_NEON) aom_sad8x16_avg = aom_sad8x16_avg_neon;
+    aom_sad8x16x3d = aom_sad8x16x3d_c;
+    if (flags & HAS_NEON) aom_sad8x16x3d = aom_sad8x16x3d_neon;
+    aom_sad8x16x4d = aom_sad8x16x4d_c;
+    if (flags & HAS_NEON) aom_sad8x16x4d = aom_sad8x16x4d_neon;
+    aom_sad8x32 = aom_sad8x32_c;
+    if (flags & HAS_NEON) aom_sad8x32 = aom_sad8x32_neon;
+    aom_sad8x32_avg = aom_sad8x32_avg_c;
+    if (flags & HAS_NEON) aom_sad8x32_avg = aom_sad8x32_avg_neon;
+    aom_sad8x32x3d = aom_sad8x32x3d_c;
+    if (flags & HAS_NEON) aom_sad8x32x3d = aom_sad8x32x3d_neon;
+    aom_sad8x32x4d = aom_sad8x32x4d_c;
+    if (flags & HAS_NEON) aom_sad8x32x4d = aom_sad8x32x4d_neon;
+    aom_sad8x4 = aom_sad8x4_c;
+    if (flags & HAS_NEON) aom_sad8x4 = aom_sad8x4_neon;
+    aom_sad8x4_avg = aom_sad8x4_avg_c;
+    if (flags & HAS_NEON) aom_sad8x4_avg = aom_sad8x4_avg_neon;
+    aom_sad8x4x3d = aom_sad8x4x3d_c;
+    if (flags & HAS_NEON) aom_sad8x4x3d = aom_sad8x4x3d_neon;
+    aom_sad8x4x4d = aom_sad8x4x4d_c;
+    if (flags & HAS_NEON) aom_sad8x4x4d = aom_sad8x4x4d_neon;
+    aom_sad8x8 = aom_sad8x8_c;
+    if (flags & HAS_NEON) aom_sad8x8 = aom_sad8x8_neon;
+    aom_sad8x8_avg = aom_sad8x8_avg_c;
+    if (flags & HAS_NEON) aom_sad8x8_avg = aom_sad8x8_avg_neon;
+    aom_sad8x8x3d = aom_sad8x8x3d_c;
+    if (flags & HAS_NEON) aom_sad8x8x3d = aom_sad8x8x3d_neon;
+    aom_sad8x8x4d = aom_sad8x8x4d_c;
+    if (flags & HAS_NEON) aom_sad8x8x4d = aom_sad8x8x4d_neon;
+    aom_sad_skip_128x128 = aom_sad_skip_128x128_c;
+    if (flags & HAS_NEON) aom_sad_skip_128x128 = aom_sad_skip_128x128_neon;
+    aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_neon;
+    aom_sad_skip_128x64 = aom_sad_skip_128x64_c;
+    if (flags & HAS_NEON) aom_sad_skip_128x64 = aom_sad_skip_128x64_neon;
+    aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_neon;
+    aom_sad_skip_16x16 = aom_sad_skip_16x16_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x16 = aom_sad_skip_16x16_neon;
+    aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_neon;
+    aom_sad_skip_16x32 = aom_sad_skip_16x32_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x32 = aom_sad_skip_16x32_neon;
+    aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_neon;
+    aom_sad_skip_16x4 = aom_sad_skip_16x4_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x4 = aom_sad_skip_16x4_neon;
+    aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_neon;
+    aom_sad_skip_16x64 = aom_sad_skip_16x64_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x64 = aom_sad_skip_16x64_neon;
+    aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_neon;
+    aom_sad_skip_16x8 = aom_sad_skip_16x8_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x8 = aom_sad_skip_16x8_neon;
+    aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_neon;
+    aom_sad_skip_32x16 = aom_sad_skip_32x16_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x16 = aom_sad_skip_32x16_neon;
+    aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_neon;
+    aom_sad_skip_32x32 = aom_sad_skip_32x32_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x32 = aom_sad_skip_32x32_neon;
+    aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_neon;
+    aom_sad_skip_32x64 = aom_sad_skip_32x64_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x64 = aom_sad_skip_32x64_neon;
+    aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_neon;
+    aom_sad_skip_32x8 = aom_sad_skip_32x8_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x8 = aom_sad_skip_32x8_neon;
+    aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_neon;
+    aom_sad_skip_4x16 = aom_sad_skip_4x16_c;
+    if (flags & HAS_NEON) aom_sad_skip_4x16 = aom_sad_skip_4x16_neon;
+    aom_sad_skip_4x16x4d = aom_sad_skip_4x16x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_4x16x4d = aom_sad_skip_4x16x4d_neon;
+    aom_sad_skip_4x4 = aom_sad_skip_4x4_c;
+    if (flags & HAS_NEON) aom_sad_skip_4x4 = aom_sad_skip_4x4_neon;
+    aom_sad_skip_4x4x4d = aom_sad_skip_4x4x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_4x4x4d = aom_sad_skip_4x4x4d_neon;
+    aom_sad_skip_4x8 = aom_sad_skip_4x8_c;
+    if (flags & HAS_NEON) aom_sad_skip_4x8 = aom_sad_skip_4x8_neon;
+    aom_sad_skip_4x8x4d = aom_sad_skip_4x8x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_4x8x4d = aom_sad_skip_4x8x4d_neon;
+    aom_sad_skip_64x128 = aom_sad_skip_64x128_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x128 = aom_sad_skip_64x128_neon;
+    aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_neon;
+    aom_sad_skip_64x16 = aom_sad_skip_64x16_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x16 = aom_sad_skip_64x16_neon;
+    aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_neon;
+    aom_sad_skip_64x32 = aom_sad_skip_64x32_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x32 = aom_sad_skip_64x32_neon;
+    aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_neon;
+    aom_sad_skip_64x64 = aom_sad_skip_64x64_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x64 = aom_sad_skip_64x64_neon;
+    aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_neon;
+    aom_sad_skip_8x16 = aom_sad_skip_8x16_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x16 = aom_sad_skip_8x16_neon;
+    aom_sad_skip_8x16x4d = aom_sad_skip_8x16x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x16x4d = aom_sad_skip_8x16x4d_neon;
+    aom_sad_skip_8x32 = aom_sad_skip_8x32_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x32 = aom_sad_skip_8x32_neon;
+    aom_sad_skip_8x32x4d = aom_sad_skip_8x32x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x32x4d = aom_sad_skip_8x32x4d_neon;
+    aom_sad_skip_8x4 = aom_sad_skip_8x4_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x4 = aom_sad_skip_8x4_neon;
+    aom_sad_skip_8x4x4d = aom_sad_skip_8x4x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x4x4d = aom_sad_skip_8x4x4d_neon;
+    aom_sad_skip_8x8 = aom_sad_skip_8x8_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x8 = aom_sad_skip_8x8_neon;
+    aom_sad_skip_8x8x4d = aom_sad_skip_8x8x4d_c;
+    if (flags & HAS_NEON) aom_sad_skip_8x8x4d = aom_sad_skip_8x8x4d_neon;
+    aom_satd = aom_satd_c;
+    if (flags & HAS_NEON) aom_satd = aom_satd_neon;
+    aom_satd_lp = aom_satd_lp_c;
+    if (flags & HAS_NEON) aom_satd_lp = aom_satd_lp_neon;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_NEON) aom_scaled_2d = aom_scaled_2d_neon;
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_neon;
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_neon;
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_neon;
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_neon;
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_neon;
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_neon;
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_neon;
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_neon;
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_neon;
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_neon;
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_neon;
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_neon;
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_neon;
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_neon;
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_neon;
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_neon;
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_neon;
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_neon;
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_neon;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_neon;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_neon;
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_neon;
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_neon;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_neon;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_neon;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_neon;
+    aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_neon;
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_neon;
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_neon;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_neon;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_neon;
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_neon;
+    aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_neon;
+    aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_neon;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_neon;
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_neon;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_neon;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_neon;
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_neon;
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_neon;
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_neon;
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_neon;
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_neon;
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_neon;
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_neon;
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_neon;
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_neon;
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_neon;
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_neon;
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_neon;
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_neon;
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_neon;
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_neon;
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_neon;
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_neon;
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_neon;
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_neon;
+    aom_sse = aom_sse_c;
+    if (flags & HAS_NEON) aom_sse = aom_sse_neon;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_neon;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_neon;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_neon;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_neon;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_neon;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_neon;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_neon;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_neon;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_neon;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_neon;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_neon;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_neon;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_neon;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_neon;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_neon;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_neon;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_neon;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_neon;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_neon;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_neon;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_neon;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_neon;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_neon;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_neon;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_neon;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_neon;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_neon;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_neon;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_neon;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_neon;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_neon;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_neon;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_neon;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_neon;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_neon;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_neon;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_neon;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_neon;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_neon;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_neon;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_neon;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_neon;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_neon;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_neon;
+    aom_subtract_block = aom_subtract_block_c;
+    if (flags & HAS_NEON) aom_subtract_block = aom_subtract_block_neon;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_c;
+    if (flags & HAS_NEON) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_neon;
+    aom_sum_squares_i16 = aom_sum_squares_i16_c;
+    if (flags & HAS_NEON) aom_sum_squares_i16 = aom_sum_squares_i16_neon;
+    aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_c;
+    if (flags & HAS_NEON) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_neon;
+    aom_v_predictor_16x16 = aom_v_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_v_predictor_16x16 = aom_v_predictor_16x16_neon;
+    aom_v_predictor_16x32 = aom_v_predictor_16x32_c;
+    if (flags & HAS_NEON) aom_v_predictor_16x32 = aom_v_predictor_16x32_neon;
+    aom_v_predictor_16x4 = aom_v_predictor_16x4_c;
+    if (flags & HAS_NEON) aom_v_predictor_16x4 = aom_v_predictor_16x4_neon;
+    aom_v_predictor_16x64 = aom_v_predictor_16x64_c;
+    if (flags & HAS_NEON) aom_v_predictor_16x64 = aom_v_predictor_16x64_neon;
+    aom_v_predictor_16x8 = aom_v_predictor_16x8_c;
+    if (flags & HAS_NEON) aom_v_predictor_16x8 = aom_v_predictor_16x8_neon;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_c;
+    if (flags & HAS_NEON) aom_v_predictor_32x16 = aom_v_predictor_32x16_neon;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_v_predictor_32x32 = aom_v_predictor_32x32_neon;
+    aom_v_predictor_32x64 = aom_v_predictor_32x64_c;
+    if (flags & HAS_NEON) aom_v_predictor_32x64 = aom_v_predictor_32x64_neon;
+    aom_v_predictor_32x8 = aom_v_predictor_32x8_c;
+    if (flags & HAS_NEON) aom_v_predictor_32x8 = aom_v_predictor_32x8_neon;
+    aom_v_predictor_4x16 = aom_v_predictor_4x16_c;
+    if (flags & HAS_NEON) aom_v_predictor_4x16 = aom_v_predictor_4x16_neon;
+    aom_v_predictor_4x4 = aom_v_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_v_predictor_4x4 = aom_v_predictor_4x4_neon;
+    aom_v_predictor_4x8 = aom_v_predictor_4x8_c;
+    if (flags & HAS_NEON) aom_v_predictor_4x8 = aom_v_predictor_4x8_neon;
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_c;
+    if (flags & HAS_NEON) aom_v_predictor_64x16 = aom_v_predictor_64x16_neon;
+    aom_v_predictor_64x32 = aom_v_predictor_64x32_c;
+    if (flags & HAS_NEON) aom_v_predictor_64x32 = aom_v_predictor_64x32_neon;
+    aom_v_predictor_64x64 = aom_v_predictor_64x64_c;
+    if (flags & HAS_NEON) aom_v_predictor_64x64 = aom_v_predictor_64x64_neon;
+    aom_v_predictor_8x16 = aom_v_predictor_8x16_c;
+    if (flags & HAS_NEON) aom_v_predictor_8x16 = aom_v_predictor_8x16_neon;
+    aom_v_predictor_8x32 = aom_v_predictor_8x32_c;
+    if (flags & HAS_NEON) aom_v_predictor_8x32 = aom_v_predictor_8x32_neon;
+    aom_v_predictor_8x4 = aom_v_predictor_8x4_c;
+    if (flags & HAS_NEON) aom_v_predictor_8x4 = aom_v_predictor_8x4_neon;
+    aom_v_predictor_8x8 = aom_v_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_v_predictor_8x8 = aom_v_predictor_8x8_neon;
+    aom_var_2d_u16 = aom_var_2d_u16_c;
+    if (flags & HAS_NEON) aom_var_2d_u16 = aom_var_2d_u16_neon;
+    aom_var_2d_u8 = aom_var_2d_u8_c;
+    if (flags & HAS_NEON) aom_var_2d_u8 = aom_var_2d_u8_neon;
+    aom_variance128x128 = aom_variance128x128_c;
+    if (flags & HAS_NEON) aom_variance128x128 = aom_variance128x128_neon;
+    aom_variance128x64 = aom_variance128x64_c;
+    if (flags & HAS_NEON) aom_variance128x64 = aom_variance128x64_neon;
+    aom_variance16x16 = aom_variance16x16_c;
+    if (flags & HAS_NEON) aom_variance16x16 = aom_variance16x16_neon;
+    aom_variance16x32 = aom_variance16x32_c;
+    if (flags & HAS_NEON) aom_variance16x32 = aom_variance16x32_neon;
+    aom_variance16x4 = aom_variance16x4_c;
+    if (flags & HAS_NEON) aom_variance16x4 = aom_variance16x4_neon;
+    aom_variance16x64 = aom_variance16x64_c;
+    if (flags & HAS_NEON) aom_variance16x64 = aom_variance16x64_neon;
+    aom_variance16x8 = aom_variance16x8_c;
+    if (flags & HAS_NEON) aom_variance16x8 = aom_variance16x8_neon;
+    aom_variance32x16 = aom_variance32x16_c;
+    if (flags & HAS_NEON) aom_variance32x16 = aom_variance32x16_neon;
+    aom_variance32x32 = aom_variance32x32_c;
+    if (flags & HAS_NEON) aom_variance32x32 = aom_variance32x32_neon;
+    aom_variance32x64 = aom_variance32x64_c;
+    if (flags & HAS_NEON) aom_variance32x64 = aom_variance32x64_neon;
+    aom_variance32x8 = aom_variance32x8_c;
+    if (flags & HAS_NEON) aom_variance32x8 = aom_variance32x8_neon;
+    aom_variance4x16 = aom_variance4x16_c;
+    if (flags & HAS_NEON) aom_variance4x16 = aom_variance4x16_neon;
+    aom_variance4x4 = aom_variance4x4_c;
+    if (flags & HAS_NEON) aom_variance4x4 = aom_variance4x4_neon;
+    aom_variance4x8 = aom_variance4x8_c;
+    if (flags & HAS_NEON) aom_variance4x8 = aom_variance4x8_neon;
+    aom_variance64x128 = aom_variance64x128_c;
+    if (flags & HAS_NEON) aom_variance64x128 = aom_variance64x128_neon;
+    aom_variance64x16 = aom_variance64x16_c;
+    if (flags & HAS_NEON) aom_variance64x16 = aom_variance64x16_neon;
+    aom_variance64x32 = aom_variance64x32_c;
+    if (flags & HAS_NEON) aom_variance64x32 = aom_variance64x32_neon;
+    aom_variance64x64 = aom_variance64x64_c;
+    if (flags & HAS_NEON) aom_variance64x64 = aom_variance64x64_neon;
+    aom_variance8x16 = aom_variance8x16_c;
+    if (flags & HAS_NEON) aom_variance8x16 = aom_variance8x16_neon;
+    aom_variance8x32 = aom_variance8x32_c;
+    if (flags & HAS_NEON) aom_variance8x32 = aom_variance8x32_neon;
+    aom_variance8x4 = aom_variance8x4_c;
+    if (flags & HAS_NEON) aom_variance8x4 = aom_variance8x4_neon;
+    aom_variance8x8 = aom_variance8x8_c;
+    if (flags & HAS_NEON) aom_variance8x8 = aom_variance8x8_neon;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_NEON) aom_vector_var = aom_vector_var_neon;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/arm/config/aom_scale_rtcd.h b/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..d296957f84
--- /dev/null
+++ b/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
@@ -0,0 +1,105 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#include "config/aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = aom_arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/arm/config/av1_rtcd.h b/media/libaom/config/linux/arm/config/av1_rtcd.h
new file mode 100644
index 0000000000..958400d252
--- /dev/null
+++ b/media/libaom/config/linux/arm/config/av1_rtcd.h
@@ -0,0 +1,1145 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+void aom_comp_avg_upsampled_pred_neon(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+RTCD_EXTERN void (*aom_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_dist_wtd_comp_avg_upsampled_pred_neon(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_comp_avg_upsampled_pred_neon(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_neon(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_upsampled_pred_neon(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+void aom_quantize_b_helper_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+RTCD_EXTERN void (*aom_quantize_b_helper)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+void aom_upsampled_pred_neon(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+RTCD_EXTERN void (*aom_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_neon(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_neon(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_neon(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+RTCD_EXTERN int64_t (*av1_block_error_lp)(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_neon(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_neon(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_neon(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_neon(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim1)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_neon(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim2)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_neon(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_neon(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params_high_bd)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+void av1_cnn_convolve_no_maxpool_padding_valid_neon(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+RTCD_EXTERN void (*av1_cnn_convolve_no_maxpool_padding_valid)(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_neon(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+RTCD_EXTERN void (*av1_compute_stats)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_neon(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+RTCD_EXTERN void (*av1_compute_stats_highbd)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_scale av1_convolve_2d_scale_c
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_intrabc_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr_intrabc)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+#define av1_convolve_horiz_rs av1_convolve_horiz_rs_c
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_intrabc_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr_intrabc)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_intrabc_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr_intrabc)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_neon(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_neon(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_neon(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+double av1_estimate_noise_from_single_plane_neon(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+RTCD_EXTERN double (*av1_estimate_noise_from_single_plane)(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_neon(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+void av1_filter_intra_predictor_neon(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x32_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x4_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x64_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x16_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x64_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x8_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x16_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x8_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x16_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x32_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x32_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x4_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_neon(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void av1_fwht4x4_neon(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*av1_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+#define av1_get_crc32c_value av1_get_crc32c_value_c
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_neon(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+RTCD_EXTERN void (*av1_get_horver_correlation_full)(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+void av1_get_nz_map_contexts_neon(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+RTCD_EXTERN void (*av1_get_nz_map_contexts)(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_neon(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_highbd_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+RTCD_EXTERN int64_t (*av1_highbd_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_c
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_c
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_intrabc_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr_intrabc)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+void av1_highbd_convolve_horiz_rs_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_intrabc_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr_intrabc)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_intrabc_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr_intrabc)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_neon(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z1_neon(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z2_neon(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z3_neon(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+double av1_highbd_estimate_noise_from_single_plane_neon(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+RTCD_EXTERN double (*av1_highbd_estimate_noise_from_single_plane)(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+void av1_highbd_filter_intra_edge_neon(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_highbd_filter_intra_edge)(uint16_t *p, int sz, int strength);
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x32_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x32)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x4_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x8_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x16_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x64_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x64)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x16_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x8_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_64x32_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_64x32)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_64x64_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_64x64)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x16_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x4_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_neon(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_highbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+RTCD_EXTERN void (*av1_highbd_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+void av1_highbd_upsample_intra_edge_neon(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_highbd_upsample_intra_edge)(uint16_t *p, int sz, int bd);
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_neon(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_16x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_16x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_16x4_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_16x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_16x64_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_16x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_16x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_16x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_32x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_32x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_32x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_32x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_32x64_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_32x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_32x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_32x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_64x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_64x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_64x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_64x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_64x64_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_64x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x4_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_neon(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_neon(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_neon(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_lowbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+#define av1_nn_fast_softmax_16 av1_nn_fast_softmax_16_c
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_neon(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+RTCD_EXTERN void (*av1_nn_predict)(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_64x64_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_neon(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_lp)(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+void av1_resize_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+void av1_round_shift_array_neon(int32_t *arr, int size, int bit);
+RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit);
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_neon(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_neon(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+RTCD_EXTERN void (*av1_txb_init_levels)(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_neon(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_neon(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_neon(int16_t *d, const int16_t *a, const int16_t *b, int N);
+RTCD_EXTERN void (*av1_wedge_compute_delta_squares)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_neon(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+RTCD_EXTERN int8_t (*av1_wedge_sign_from_residuals)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_neon(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+RTCD_EXTERN uint64_t (*av1_wedge_sse_from_residuals)(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_neon(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_neon(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_neon(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_neon(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*cfl_get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*cfl_get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_neon(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*cfl_get_subtract_average_fn)(TX_SIZE tx_size);
+
+void av1_rtcd(void);
+
+#include "config/aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = aom_arm_cpu_caps();
+
+    (void)flags;
+
+    aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_c;
+    if (flags & HAS_NEON) aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_neon;
+    aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_NEON) aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_neon;
+    aom_highbd_comp_avg_upsampled_pred = aom_highbd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_NEON) aom_highbd_comp_avg_upsampled_pred = aom_highbd_comp_avg_upsampled_pred_neon;
+    aom_highbd_dist_wtd_comp_avg_upsampled_pred = aom_highbd_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_NEON) aom_highbd_dist_wtd_comp_avg_upsampled_pred = aom_highbd_dist_wtd_comp_avg_upsampled_pred_neon;
+    aom_highbd_upsampled_pred = aom_highbd_upsampled_pred_c;
+    if (flags & HAS_NEON) aom_highbd_upsampled_pred = aom_highbd_upsampled_pred_neon;
+    aom_quantize_b_helper = aom_quantize_b_helper_c;
+    if (flags & HAS_NEON) aom_quantize_b_helper = aom_quantize_b_helper_neon;
+    aom_upsampled_pred = aom_upsampled_pred_c;
+    if (flags & HAS_NEON) aom_upsampled_pred = aom_upsampled_pred_neon;
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+    if (flags & HAS_NEON) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_neon;
+    av1_apply_temporal_filter = av1_apply_temporal_filter_c;
+    if (flags & HAS_NEON) av1_apply_temporal_filter = av1_apply_temporal_filter_neon;
+    av1_block_error = av1_block_error_c;
+    if (flags & HAS_NEON) av1_block_error = av1_block_error_neon;
+    av1_block_error_lp = av1_block_error_lp_c;
+    if (flags & HAS_NEON) av1_block_error_lp = av1_block_error_lp_neon;
+    av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
+    if (flags & HAS_NEON) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_neon;
+    av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
+    if (flags & HAS_NEON) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_neon;
+    av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
+    if (flags & HAS_NEON) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_neon;
+    av1_calc_indices_dim1 = av1_calc_indices_dim1_c;
+    if (flags & HAS_NEON) av1_calc_indices_dim1 = av1_calc_indices_dim1_neon;
+    av1_calc_indices_dim2 = av1_calc_indices_dim2_c;
+    if (flags & HAS_NEON) av1_calc_indices_dim2 = av1_calc_indices_dim2_neon;
+    av1_calc_proj_params = av1_calc_proj_params_c;
+    if (flags & HAS_NEON) av1_calc_proj_params = av1_calc_proj_params_neon;
+    av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_c;
+    if (flags & HAS_NEON) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_neon;
+    av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_c;
+    if (flags & HAS_NEON) av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_neon;
+    av1_compute_stats = av1_compute_stats_c;
+    if (flags & HAS_NEON) av1_compute_stats = av1_compute_stats_neon;
+    av1_compute_stats_highbd = av1_compute_stats_highbd_c;
+    if (flags & HAS_NEON) av1_compute_stats_highbd = av1_compute_stats_highbd_neon;
+    av1_convolve_2d_sr = av1_convolve_2d_sr_c;
+    if (flags & HAS_NEON) av1_convolve_2d_sr = av1_convolve_2d_sr_neon;
+    av1_convolve_2d_sr_intrabc = av1_convolve_2d_sr_intrabc_c;
+    if (flags & HAS_NEON) av1_convolve_2d_sr_intrabc = av1_convolve_2d_sr_intrabc_neon;
+    av1_convolve_x_sr = av1_convolve_x_sr_c;
+    if (flags & HAS_NEON) av1_convolve_x_sr = av1_convolve_x_sr_neon;
+    av1_convolve_x_sr_intrabc = av1_convolve_x_sr_intrabc_c;
+    if (flags & HAS_NEON) av1_convolve_x_sr_intrabc = av1_convolve_x_sr_intrabc_neon;
+    av1_convolve_y_sr = av1_convolve_y_sr_c;
+    if (flags & HAS_NEON) av1_convolve_y_sr = av1_convolve_y_sr_neon;
+    av1_convolve_y_sr_intrabc = av1_convolve_y_sr_intrabc_c;
+    if (flags & HAS_NEON) av1_convolve_y_sr_intrabc = av1_convolve_y_sr_intrabc_neon;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
+    if (flags & HAS_NEON) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_neon;
+    av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_NEON) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_neon;
+    av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_c;
+    if (flags & HAS_NEON) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_neon;
+    av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_c;
+    if (flags & HAS_NEON) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_neon;
+    av1_dr_prediction_z1 = av1_dr_prediction_z1_c;
+    if (flags & HAS_NEON) av1_dr_prediction_z1 = av1_dr_prediction_z1_neon;
+    av1_dr_prediction_z2 = av1_dr_prediction_z2_c;
+    if (flags & HAS_NEON) av1_dr_prediction_z2 = av1_dr_prediction_z2_neon;
+    av1_dr_prediction_z3 = av1_dr_prediction_z3_c;
+    if (flags & HAS_NEON) av1_dr_prediction_z3 = av1_dr_prediction_z3_neon;
+    av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_c;
+    if (flags & HAS_NEON) av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_neon;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_NEON) av1_filter_intra_edge = av1_filter_intra_edge_neon;
+    av1_filter_intra_predictor = av1_filter_intra_predictor_c;
+    if (flags & HAS_NEON) av1_filter_intra_predictor = av1_filter_intra_predictor_neon;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_neon;
+    av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_neon;
+    av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_neon;
+    av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_neon;
+    av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_neon;
+    av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_neon;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_neon;
+    av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_neon;
+    av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_neon;
+    av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_neon;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_neon;
+    av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_neon;
+    av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_neon;
+    av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_neon;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_neon;
+    av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_neon;
+    av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_neon;
+    av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_neon;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_NEON) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_neon;
+    av1_fwht4x4 = av1_fwht4x4_c;
+    if (flags & HAS_NEON) av1_fwht4x4 = av1_fwht4x4_neon;
+    av1_get_horver_correlation_full = av1_get_horver_correlation_full_c;
+    if (flags & HAS_NEON) av1_get_horver_correlation_full = av1_get_horver_correlation_full_neon;
+    av1_get_nz_map_contexts = av1_get_nz_map_contexts_c;
+    if (flags & HAS_NEON) av1_get_nz_map_contexts = av1_get_nz_map_contexts_neon;
+    av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_c;
+    if (flags & HAS_NEON) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_neon;
+    av1_highbd_block_error = av1_highbd_block_error_c;
+    if (flags & HAS_NEON) av1_highbd_block_error = av1_highbd_block_error_neon;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_neon;
+    av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_neon;
+    av1_highbd_convolve_2d_sr_intrabc = av1_highbd_convolve_2d_sr_intrabc_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_2d_sr_intrabc = av1_highbd_convolve_2d_sr_intrabc_neon;
+    av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_neon;
+    av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_neon;
+    av1_highbd_convolve_x_sr_intrabc = av1_highbd_convolve_x_sr_intrabc_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_x_sr_intrabc = av1_highbd_convolve_x_sr_intrabc_neon;
+    av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_neon;
+    av1_highbd_convolve_y_sr_intrabc = av1_highbd_convolve_y_sr_intrabc_c;
+    if (flags & HAS_NEON) av1_highbd_convolve_y_sr_intrabc = av1_highbd_convolve_y_sr_intrabc_neon;
+    av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_c;
+    if (flags & HAS_NEON) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_neon;
+    av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_NEON) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_neon;
+    av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_c;
+    if (flags & HAS_NEON) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_neon;
+    av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_c;
+    if (flags & HAS_NEON) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_neon;
+    av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_c;
+    if (flags & HAS_NEON) av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_neon;
+    av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_c;
+    if (flags & HAS_NEON) av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_neon;
+    av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_c;
+    if (flags & HAS_NEON) av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_neon;
+    av1_highbd_estimate_noise_from_single_plane = av1_highbd_estimate_noise_from_single_plane_c;
+    if (flags & HAS_NEON) av1_highbd_estimate_noise_from_single_plane = av1_highbd_estimate_noise_from_single_plane_neon;
+    av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_c;
+    if (flags & HAS_NEON) av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_neon;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_neon;
+    av1_highbd_inv_txfm_add_16x32 = av1_highbd_inv_txfm_add_16x32_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_16x32 = av1_highbd_inv_txfm_add_16x32_neon;
+    av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_neon;
+    av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_neon;
+    av1_highbd_inv_txfm_add_32x16 = av1_highbd_inv_txfm_add_32x16_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_32x16 = av1_highbd_inv_txfm_add_32x16_neon;
+    av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_neon;
+    av1_highbd_inv_txfm_add_32x64 = av1_highbd_inv_txfm_add_32x64_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_32x64 = av1_highbd_inv_txfm_add_32x64_neon;
+    av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_neon;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_neon;
+    av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_neon;
+    av1_highbd_inv_txfm_add_64x32 = av1_highbd_inv_txfm_add_64x32_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_64x32 = av1_highbd_inv_txfm_add_64x32_neon;
+    av1_highbd_inv_txfm_add_64x64 = av1_highbd_inv_txfm_add_64x64_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_64x64 = av1_highbd_inv_txfm_add_64x64_neon;
+    av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_neon;
+    av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_neon;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_NEON) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_neon;
+    av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_c;
+    if (flags & HAS_NEON) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_neon;
+    av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
+    if (flags & HAS_NEON) av1_highbd_quantize_fp = av1_highbd_quantize_fp_neon;
+    av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_c;
+    if (flags & HAS_NEON) av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_neon;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_NEON) av1_highbd_warp_affine = av1_highbd_warp_affine_neon;
+    av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
+    if (flags & HAS_NEON) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_neon;
+    av1_inv_txfm2d_add_16x32 = av1_inv_txfm2d_add_16x32_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_16x32 = av1_inv_txfm2d_add_16x32_neon;
+    av1_inv_txfm2d_add_16x4 = av1_inv_txfm2d_add_16x4_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_16x4 = av1_inv_txfm2d_add_16x4_neon;
+    av1_inv_txfm2d_add_16x64 = av1_inv_txfm2d_add_16x64_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_16x64 = av1_inv_txfm2d_add_16x64_neon;
+    av1_inv_txfm2d_add_16x8 = av1_inv_txfm2d_add_16x8_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_16x8 = av1_inv_txfm2d_add_16x8_neon;
+    av1_inv_txfm2d_add_32x16 = av1_inv_txfm2d_add_32x16_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_32x16 = av1_inv_txfm2d_add_32x16_neon;
+    av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_neon;
+    av1_inv_txfm2d_add_32x64 = av1_inv_txfm2d_add_32x64_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_32x64 = av1_inv_txfm2d_add_32x64_neon;
+    av1_inv_txfm2d_add_32x8 = av1_inv_txfm2d_add_32x8_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_32x8 = av1_inv_txfm2d_add_32x8_neon;
+    av1_inv_txfm2d_add_4x16 = av1_inv_txfm2d_add_4x16_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_4x16 = av1_inv_txfm2d_add_4x16_neon;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_neon;
+    av1_inv_txfm2d_add_4x8 = av1_inv_txfm2d_add_4x8_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_4x8 = av1_inv_txfm2d_add_4x8_neon;
+    av1_inv_txfm2d_add_64x16 = av1_inv_txfm2d_add_64x16_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_64x16 = av1_inv_txfm2d_add_64x16_neon;
+    av1_inv_txfm2d_add_64x32 = av1_inv_txfm2d_add_64x32_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_64x32 = av1_inv_txfm2d_add_64x32_neon;
+    av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_neon;
+    av1_inv_txfm2d_add_8x16 = av1_inv_txfm2d_add_8x16_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_8x16 = av1_inv_txfm2d_add_8x16_neon;
+    av1_inv_txfm2d_add_8x32 = av1_inv_txfm2d_add_8x32_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_8x32 = av1_inv_txfm2d_add_8x32_neon;
+    av1_inv_txfm2d_add_8x4 = av1_inv_txfm2d_add_8x4_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_8x4 = av1_inv_txfm2d_add_8x4_neon;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    if (flags & HAS_NEON) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_neon;
+    av1_inv_txfm_add = av1_inv_txfm_add_c;
+    if (flags & HAS_NEON) av1_inv_txfm_add = av1_inv_txfm_add_neon;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_c;
+    if (flags & HAS_NEON) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_neon;
+    av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
+    if (flags & HAS_NEON) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_neon;
+    av1_nn_predict = av1_nn_predict_c;
+    if (flags & HAS_NEON) av1_nn_predict = av1_nn_predict_neon;
+    av1_quantize_fp = av1_quantize_fp_c;
+    if (flags & HAS_NEON) av1_quantize_fp = av1_quantize_fp_neon;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_NEON) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_neon;
+    av1_quantize_fp_64x64 = av1_quantize_fp_64x64_c;
+    if (flags & HAS_NEON) av1_quantize_fp_64x64 = av1_quantize_fp_64x64_neon;
+    av1_quantize_lp = av1_quantize_lp_c;
+    if (flags & HAS_NEON) av1_quantize_lp = av1_quantize_lp_neon;
+    av1_resize_and_extend_frame = av1_resize_and_extend_frame_c;
+    if (flags & HAS_NEON) av1_resize_and_extend_frame = av1_resize_and_extend_frame_neon;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_NEON) av1_round_shift_array = av1_round_shift_array_neon;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_NEON) av1_selfguided_restoration = av1_selfguided_restoration_neon;
+    av1_txb_init_levels = av1_txb_init_levels_c;
+    if (flags & HAS_NEON) av1_txb_init_levels = av1_txb_init_levels_neon;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_NEON) av1_upsample_intra_edge = av1_upsample_intra_edge_neon;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_NEON) av1_warp_affine = av1_warp_affine_neon;
+    av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_c;
+    if (flags & HAS_NEON) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_neon;
+    av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_c;
+    if (flags & HAS_NEON) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_neon;
+    av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_c;
+    if (flags & HAS_NEON) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_neon;
+    av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_c;
+    if (flags & HAS_NEON) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_neon;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
+    if (flags & HAS_NEON) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_neon;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
+    if (flags & HAS_NEON) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_neon;
+    cdef_filter_16_0 = cdef_filter_16_0_c;
+    if (flags & HAS_NEON) cdef_filter_16_0 = cdef_filter_16_0_neon;
+    cdef_filter_16_1 = cdef_filter_16_1_c;
+    if (flags & HAS_NEON) cdef_filter_16_1 = cdef_filter_16_1_neon;
+    cdef_filter_16_2 = cdef_filter_16_2_c;
+    if (flags & HAS_NEON) cdef_filter_16_2 = cdef_filter_16_2_neon;
+    cdef_filter_16_3 = cdef_filter_16_3_c;
+    if (flags & HAS_NEON) cdef_filter_16_3 = cdef_filter_16_3_neon;
+    cdef_filter_8_0 = cdef_filter_8_0_c;
+    if (flags & HAS_NEON) cdef_filter_8_0 = cdef_filter_8_0_neon;
+    cdef_filter_8_1 = cdef_filter_8_1_c;
+    if (flags & HAS_NEON) cdef_filter_8_1 = cdef_filter_8_1_neon;
+    cdef_filter_8_2 = cdef_filter_8_2_c;
+    if (flags & HAS_NEON) cdef_filter_8_2 = cdef_filter_8_2_neon;
+    cdef_filter_8_3 = cdef_filter_8_3_c;
+    if (flags & HAS_NEON) cdef_filter_8_3 = cdef_filter_8_3_neon;
+    cdef_find_dir = cdef_find_dir_c;
+    if (flags & HAS_NEON) cdef_find_dir = cdef_find_dir_neon;
+    cdef_find_dir_dual = cdef_find_dir_dual_c;
+    if (flags & HAS_NEON) cdef_find_dir_dual = cdef_find_dir_dual_neon;
+    cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
+    if (flags & HAS_NEON) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_neon;
+    cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
+    if (flags & HAS_NEON) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_neon;
+    cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
+    if (flags & HAS_NEON) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_neon;
+    cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
+    if (flags & HAS_NEON) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_neon;
+    cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
+    if (flags & HAS_NEON) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_neon;
+    cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
+    if (flags & HAS_NEON) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_neon;
+    cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_c;
+    if (flags & HAS_NEON) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_neon;
+    cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_c;
+    if (flags & HAS_NEON) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_neon;
+    cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_c;
+    if (flags & HAS_NEON) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_neon;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/ia32/config/aom_config.asm b/media/libaom/config/linux/ia32/config/aom_config.asm
new file mode 100644
index 0000000000..e75260cb09
--- /dev/null
+++ b/media/libaom/config/linux/ia32/config/aom_config.asm
@@ -0,0 +1,92 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+AOM_ARCH_AARCH64 equ 0
+AOM_ARCH_ARM equ 0
+AOM_ARCH_PPC equ 0
+AOM_ARCH_X86 equ 1
+AOM_ARCH_X86_64 equ 0
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_HIGHBITDEPTH equ 1
+CONFIG_AV1_TEMPORAL_DENOISING equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITRATE_ACCURACY equ 0
+CONFIG_BITRATE_ACCURACY_BL equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_COMPONENT_TIMING equ 0
+CONFIG_COLLECT_PARTITION_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_CWG_C013 equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_EXCLUDE_SIMD_MISMATCH equ 0
+CONFIG_FPMT_TEST equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBVMAF_PSNR_PEAK equ 1
+CONFIG_LIBYUV equ 0
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NN_V2 equ 0
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OPTICAL_FLOW_API equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_OUTPUT_FRAME_SIZE equ 0
+CONFIG_PARTITION_SEARCH_ORDER equ 0
+CONFIG_PIC equ 1
+CONFIG_RATECTRL_LOG equ 0
+CONFIG_RD_COMMAND equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_RT_ML_PARTITIONING equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SALIENCY_MAP equ 0
+CONFIG_SHARED equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_SPEED_STATS equ 0
+CONFIG_TFLITE equ 0
+CONFIG_THREE_PASS equ 0
+CONFIG_TUNE_BUTTERAUGLI equ 0
+CONFIG_TUNE_VMAF equ 0
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+FORCE_HIGHBITDEPTH_DECODING equ 0
+HAVE_ARM_CRC32 equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_FEXCEPT equ 1
+HAVE_MMX equ 1
+HAVE_NEON equ 0
+HAVE_NEON_DOTPROD equ 0
+HAVE_NEON_I8MM equ 0
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 1
+HAVE_SSE4_1 equ 1
+HAVE_SSE4_2 equ 1
+HAVE_SSSE3 equ 1
+HAVE_SVE equ 0
+HAVE_VSX equ 0
+HAVE_WXWIDGETS equ 0
+STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/linux/ia32/config/aom_config.h b/media/libaom/config/linux/ia32/config/aom_config.h
new file mode 100644
index 0000000000..b0e5b5cabc
--- /dev/null
+++ b/media/libaom/config/linux/ia32/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 0
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 1
+#define AOM_ARCH_X86_64 0
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 1
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 1
+#define HAVE_NEON 0
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_SSE4_2 1
+#define HAVE_SSSE3 1
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..a19adf5f61
--- /dev/null
+++ b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
@@ -0,0 +1,9185 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_4x4)(const uint8_t *, int p);
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_8x8)(const uint8_t *, int p);
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_avx2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_dist_wtd_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+void aom_fdct4x4_lp_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4_lp)(const int16_t *input, int16_t *output, int stride);
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+void aom_fft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_fft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+void aom_fft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_fft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+void aom_fft4x4_float_sse2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft4x4_float)(const float *input, float *temp, float *output);
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+void aom_fft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_fft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_sse2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_avx2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+RTCD_EXTERN void (*aom_get_blk_sse_sum)(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+RTCD_EXTERN unsigned int (*aom_get_mb_ss)(const int16_t *);
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+RTCD_EXTERN void (*aom_get_var_sse_sum_16x16_dual)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+RTCD_EXTERN void (*aom_get_var_sse_sum_8x8_quad)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_4x4)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8_dual)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x128 aom_highbd_10_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x64 aom_highbd_10_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x64 aom_highbd_10_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x128 aom_highbd_10_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x16 aom_highbd_10_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x128 aom_highbd_10_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x64 aom_highbd_10_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x128 aom_highbd_10_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x128 aom_highbd_12_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x64 aom_highbd_12_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x64 aom_highbd_12_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x128 aom_highbd_12_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x16 aom_highbd_12_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x128 aom_highbd_12_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x64 aom_highbd_12_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x128 aom_highbd_12_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x128 aom_highbd_8_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x64 aom_highbd_8_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x16 aom_highbd_8_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x32 aom_highbd_8_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x4 aom_highbd_8_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x64 aom_highbd_8_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x8 aom_highbd_8_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x16 aom_highbd_8_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x32 aom_highbd_8_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x64 aom_highbd_8_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x8 aom_highbd_8_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x16 aom_highbd_8_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x4 aom_highbd_8_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x8 aom_highbd_8_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x128 aom_highbd_8_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x16 aom_highbd_8_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x32 aom_highbd_8_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x64 aom_highbd_8_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x16 aom_highbd_8_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x32 aom_highbd_8_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x4 aom_highbd_8_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x8 aom_highbd_8_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x128 aom_highbd_8_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x64 aom_highbd_8_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x128 aom_highbd_8_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+#define aom_highbd_avg_4x4 aom_highbd_avg_4x4_c
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+#define aom_highbd_avg_8x8 aom_highbd_avg_8x8_c
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+#define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_highbd_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x4 aom_highbd_dc_128_predictor_16x4_c
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x64 aom_highbd_dc_128_predictor_16x64_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x64 aom_highbd_dc_128_predictor_32x64_c
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x8 aom_highbd_dc_128_predictor_32x8_c
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x16 aom_highbd_dc_128_predictor_4x16_c
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x16 aom_highbd_dc_128_predictor_64x16_c
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x32 aom_highbd_dc_128_predictor_64x32_c
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x64 aom_highbd_dc_128_predictor_64x64_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x32 aom_highbd_dc_128_predictor_8x32_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x4 aom_highbd_dc_left_predictor_16x4_c
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x64 aom_highbd_dc_left_predictor_16x64_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x64 aom_highbd_dc_left_predictor_32x64_c
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x8 aom_highbd_dc_left_predictor_32x8_c
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x16 aom_highbd_dc_left_predictor_4x16_c
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x16 aom_highbd_dc_left_predictor_64x16_c
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x32 aom_highbd_dc_left_predictor_64x32_c
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x64 aom_highbd_dc_left_predictor_64x64_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x32 aom_highbd_dc_left_predictor_8x32_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x4 aom_highbd_dc_predictor_16x4_c
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x64 aom_highbd_dc_predictor_16x64_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x64 aom_highbd_dc_predictor_32x64_c
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x8 aom_highbd_dc_predictor_32x8_c
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x16 aom_highbd_dc_predictor_4x16_c
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x16 aom_highbd_dc_predictor_64x16_c
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x32 aom_highbd_dc_predictor_64x32_c
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x64 aom_highbd_dc_predictor_64x64_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x32 aom_highbd_dc_predictor_8x32_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x4 aom_highbd_dc_top_predictor_16x4_c
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x64 aom_highbd_dc_top_predictor_16x64_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x64 aom_highbd_dc_top_predictor_32x64_c
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x8 aom_highbd_dc_top_predictor_32x8_c
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x16 aom_highbd_dc_top_predictor_4x16_c
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x16 aom_highbd_dc_top_predictor_64x16_c
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x32 aom_highbd_dc_top_predictor_64x32_c
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x64 aom_highbd_dc_top_predictor_64x64_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x32 aom_highbd_dc_top_predictor_8x32_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_highbd_dist_wtd_comp_avg_pred_sse2(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_highbd_dist_wtd_comp_avg_pred)(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x4 aom_highbd_h_predictor_16x4_c
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x64 aom_highbd_h_predictor_16x64_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x64 aom_highbd_h_predictor_32x64_c
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x8 aom_highbd_h_predictor_32x8_c
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x16 aom_highbd_h_predictor_4x16_c
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x16 aom_highbd_h_predictor_64x16_c
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x32 aom_highbd_h_predictor_64x32_c
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x64 aom_highbd_h_predictor_64x64_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x32 aom_highbd_h_predictor_8x32_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x3d aom_highbd_sad4x16x3d_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x3d aom_highbd_sad4x4x3d_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x3d aom_highbd_sad4x8x3d_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x3d aom_highbd_sad8x16x3d_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x3d aom_highbd_sad8x32x3d_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x3d aom_highbd_sad8x4x3d_c
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x3d aom_highbd_sad8x8x3d_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x4 aom_highbd_sad_skip_16x4_c
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x4x4d aom_highbd_sad_skip_16x4x4d_c
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x4 aom_highbd_sad_skip_4x4_c
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x4x4d aom_highbd_sad_skip_4x4x4d_c
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x4 aom_highbd_sad_skip_8x4_c
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x4x4d aom_highbd_sad_skip_8x4x4d_c
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_highbd_sse)(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_highbd_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft4x4_float)(const float *input, float *temp, float *output);
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_avx2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_col)(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_avx2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+RTCD_EXTERN void (*aom_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_16xh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_sse2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_avx2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit_highbd)(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x3d aom_sad16x4x3d_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x3d aom_sad4x16x3d_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x3d aom_sad4x4x3d_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x3d aom_sad4x8x3d_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x3d aom_sad8x16x3d_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x3d aom_sad8x32x3d_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x3d aom_sad8x4x3d_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x3d aom_sad8x8x3d_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x4 aom_sad_skip_16x4_c
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x4 aom_sad_skip_4x4_c
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x4x4d aom_sad_skip_4x4x4d_c
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x4 aom_sad_skip_8x4_c
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x4x4d aom_sad_skip_8x4x4d_c
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+int aom_satd_sse2(const tran_low_t *coeff, int length);
+int aom_satd_avx2(const tran_low_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const tran_low_t *coeff, int length);
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+int aom_satd_lp_sse2(const int16_t *coeff, int length);
+int aom_satd_lp_avx2(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd_lp)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_sse)(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_c
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t N);
+RTCD_EXTERN uint64_t (*aom_sum_squares_i16)(const int16_t *src, uint32_t N);
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width, int height, int *sum);
+RTCD_EXTERN uint64_t (*aom_sum_sse_2d_i16)(const int16_t *src, int src_stride, int width, int height, int *sum);
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u16)(uint8_t *src, int src_stride, int width, int height);
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u8)(uint8_t *src, int src_stride, int width, int height);
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+
+void aom_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_avg_4x4 = aom_avg_4x4_c;
+    if (flags & HAS_SSE2) aom_avg_4x4 = aom_avg_4x4_sse2;
+    aom_avg_8x8 = aom_avg_8x8_c;
+    if (flags & HAS_SSE2) aom_avg_8x8 = aom_avg_8x8_sse2;
+    aom_avg_8x8_quad = aom_avg_8x8_quad_c;
+    if (flags & HAS_SSE2) aom_avg_8x8_quad = aom_avg_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_avg_8x8_quad = aom_avg_8x8_quad_avx2;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
+    aom_comp_avg_pred = aom_comp_avg_pred_c;
+    if (flags & HAS_AVX2) aom_comp_avg_pred = aom_comp_avg_pred_avx2;
+    aom_comp_mask_pred = aom_comp_mask_pred_c;
+    if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
+    if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_flow_at_point = aom_compute_flow_at_point_c;
+    if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
+    if (flags & HAS_SSE2) aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_horiz = aom_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_horiz = aom_convolve8_horiz_avx2;
+    aom_convolve8_vert = aom_convolve8_vert_c;
+    if (flags & HAS_SSE2) aom_convolve8_vert = aom_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
+    aom_convolve_copy = aom_convolve_copy_c;
+    if (flags & HAS_SSE2) aom_convolve_copy = aom_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2;
+    aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_sse2;
+    aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_sse2;
+    aom_dc_128_predictor_16x4 = aom_dc_128_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x4 = aom_dc_128_predictor_16x4_sse2;
+    aom_dc_128_predictor_16x64 = aom_dc_128_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x64 = aom_dc_128_predictor_16x64_sse2;
+    aom_dc_128_predictor_16x8 = aom_dc_128_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x8 = aom_dc_128_predictor_16x8_sse2;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_avx2;
+    aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+    aom_dc_128_predictor_32x8 = aom_dc_128_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x8 = aom_dc_128_predictor_32x8_sse2;
+    aom_dc_128_predictor_4x16 = aom_dc_128_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x16 = aom_dc_128_predictor_4x16_sse2;
+    aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_sse2;
+    aom_dc_128_predictor_4x8 = aom_dc_128_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x8 = aom_dc_128_predictor_4x8_sse2;
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
+    aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
+    aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_avx2;
+    aom_dc_128_predictor_8x16 = aom_dc_128_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x16 = aom_dc_128_predictor_8x16_sse2;
+    aom_dc_128_predictor_8x32 = aom_dc_128_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x32 = aom_dc_128_predictor_8x32_sse2;
+    aom_dc_128_predictor_8x4 = aom_dc_128_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x4 = aom_dc_128_predictor_8x4_sse2;
+    aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_sse2;
+    aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_sse2;
+    aom_dc_left_predictor_16x32 = aom_dc_left_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x32 = aom_dc_left_predictor_16x32_sse2;
+    aom_dc_left_predictor_16x4 = aom_dc_left_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x4 = aom_dc_left_predictor_16x4_sse2;
+    aom_dc_left_predictor_16x64 = aom_dc_left_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x64 = aom_dc_left_predictor_16x64_sse2;
+    aom_dc_left_predictor_16x8 = aom_dc_left_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x8 = aom_dc_left_predictor_16x8_sse2;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_avx2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_avx2;
+    aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+    aom_dc_left_predictor_32x8 = aom_dc_left_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x8 = aom_dc_left_predictor_32x8_sse2;
+    aom_dc_left_predictor_4x16 = aom_dc_left_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x16 = aom_dc_left_predictor_4x16_sse2;
+    aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_sse2;
+    aom_dc_left_predictor_4x8 = aom_dc_left_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x8 = aom_dc_left_predictor_4x8_sse2;
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
+    aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
+    aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_avx2;
+    aom_dc_left_predictor_8x16 = aom_dc_left_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x16 = aom_dc_left_predictor_8x16_sse2;
+    aom_dc_left_predictor_8x32 = aom_dc_left_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x32 = aom_dc_left_predictor_8x32_sse2;
+    aom_dc_left_predictor_8x4 = aom_dc_left_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x4 = aom_dc_left_predictor_8x4_sse2;
+    aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_sse2;
+    aom_dc_predictor_16x16 = aom_dc_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x16 = aom_dc_predictor_16x16_sse2;
+    aom_dc_predictor_16x32 = aom_dc_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x32 = aom_dc_predictor_16x32_sse2;
+    aom_dc_predictor_16x4 = aom_dc_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x4 = aom_dc_predictor_16x4_sse2;
+    aom_dc_predictor_16x64 = aom_dc_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x64 = aom_dc_predictor_16x64_sse2;
+    aom_dc_predictor_16x8 = aom_dc_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x8 = aom_dc_predictor_16x8_sse2;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_avx2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_avx2;
+    aom_dc_predictor_32x64 = aom_dc_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+    aom_dc_predictor_32x8 = aom_dc_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x8 = aom_dc_predictor_32x8_sse2;
+    aom_dc_predictor_4x16 = aom_dc_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x16 = aom_dc_predictor_4x16_sse2;
+    aom_dc_predictor_4x4 = aom_dc_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x4 = aom_dc_predictor_4x4_sse2;
+    aom_dc_predictor_4x8 = aom_dc_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x8 = aom_dc_predictor_4x8_sse2;
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
+    aom_dc_predictor_64x32 = aom_dc_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
+    aom_dc_predictor_64x64 = aom_dc_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_avx2;
+    aom_dc_predictor_8x16 = aom_dc_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x16 = aom_dc_predictor_8x16_sse2;
+    aom_dc_predictor_8x32 = aom_dc_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x32 = aom_dc_predictor_8x32_sse2;
+    aom_dc_predictor_8x4 = aom_dc_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x4 = aom_dc_predictor_8x4_sse2;
+    aom_dc_predictor_8x8 = aom_dc_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x8 = aom_dc_predictor_8x8_sse2;
+    aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_sse2;
+    aom_dc_top_predictor_16x32 = aom_dc_top_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x32 = aom_dc_top_predictor_16x32_sse2;
+    aom_dc_top_predictor_16x4 = aom_dc_top_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x4 = aom_dc_top_predictor_16x4_sse2;
+    aom_dc_top_predictor_16x64 = aom_dc_top_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x64 = aom_dc_top_predictor_16x64_sse2;
+    aom_dc_top_predictor_16x8 = aom_dc_top_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x8 = aom_dc_top_predictor_16x8_sse2;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_avx2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_avx2;
+    aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+    aom_dc_top_predictor_32x8 = aom_dc_top_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x8 = aom_dc_top_predictor_32x8_sse2;
+    aom_dc_top_predictor_4x16 = aom_dc_top_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x16 = aom_dc_top_predictor_4x16_sse2;
+    aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_sse2;
+    aom_dc_top_predictor_4x8 = aom_dc_top_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x8 = aom_dc_top_predictor_4x8_sse2;
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
+    aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
+    aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_avx2;
+    aom_dc_top_predictor_8x16 = aom_dc_top_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x16 = aom_dc_top_predictor_8x16_sse2;
+    aom_dc_top_predictor_8x32 = aom_dc_top_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x32 = aom_dc_top_predictor_8x32_sse2;
+    aom_dc_top_predictor_8x4 = aom_dc_top_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x4 = aom_dc_top_predictor_8x4_sse2;
+    aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_sse2;
+    aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_ssse3;
+    aom_dist_wtd_sad128x128_avg = aom_dist_wtd_sad128x128_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad128x128_avg = aom_dist_wtd_sad128x128_avg_sse2;
+    aom_dist_wtd_sad128x64_avg = aom_dist_wtd_sad128x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad128x64_avg = aom_dist_wtd_sad128x64_avg_sse2;
+    aom_dist_wtd_sad16x16_avg = aom_dist_wtd_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x16_avg = aom_dist_wtd_sad16x16_avg_sse2;
+    aom_dist_wtd_sad16x32_avg = aom_dist_wtd_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x32_avg = aom_dist_wtd_sad16x32_avg_sse2;
+    aom_dist_wtd_sad16x4_avg = aom_dist_wtd_sad16x4_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x4_avg = aom_dist_wtd_sad16x4_avg_sse2;
+    aom_dist_wtd_sad16x64_avg = aom_dist_wtd_sad16x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x64_avg = aom_dist_wtd_sad16x64_avg_sse2;
+    aom_dist_wtd_sad16x8_avg = aom_dist_wtd_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x8_avg = aom_dist_wtd_sad16x8_avg_sse2;
+    aom_dist_wtd_sad32x16_avg = aom_dist_wtd_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x16_avg = aom_dist_wtd_sad32x16_avg_sse2;
+    aom_dist_wtd_sad32x32_avg = aom_dist_wtd_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x32_avg = aom_dist_wtd_sad32x32_avg_sse2;
+    aom_dist_wtd_sad32x64_avg = aom_dist_wtd_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x64_avg = aom_dist_wtd_sad32x64_avg_sse2;
+    aom_dist_wtd_sad32x8_avg = aom_dist_wtd_sad32x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x8_avg = aom_dist_wtd_sad32x8_avg_sse2;
+    aom_dist_wtd_sad4x16_avg = aom_dist_wtd_sad4x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad4x16_avg = aom_dist_wtd_sad4x16_avg_sse2;
+    aom_dist_wtd_sad4x4_avg = aom_dist_wtd_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad4x4_avg = aom_dist_wtd_sad4x4_avg_sse2;
+    aom_dist_wtd_sad4x8_avg = aom_dist_wtd_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad4x8_avg = aom_dist_wtd_sad4x8_avg_sse2;
+    aom_dist_wtd_sad64x128_avg = aom_dist_wtd_sad64x128_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x128_avg = aom_dist_wtd_sad64x128_avg_sse2;
+    aom_dist_wtd_sad64x16_avg = aom_dist_wtd_sad64x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x16_avg = aom_dist_wtd_sad64x16_avg_sse2;
+    aom_dist_wtd_sad64x32_avg = aom_dist_wtd_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x32_avg = aom_dist_wtd_sad64x32_avg_sse2;
+    aom_dist_wtd_sad64x64_avg = aom_dist_wtd_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x64_avg = aom_dist_wtd_sad64x64_avg_sse2;
+    aom_dist_wtd_sad8x16_avg = aom_dist_wtd_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x16_avg = aom_dist_wtd_sad8x16_avg_sse2;
+    aom_dist_wtd_sad8x32_avg = aom_dist_wtd_sad8x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x32_avg = aom_dist_wtd_sad8x32_avg_sse2;
+    aom_dist_wtd_sad8x4_avg = aom_dist_wtd_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x4_avg = aom_dist_wtd_sad8x4_avg_sse2;
+    aom_dist_wtd_sad8x8_avg = aom_dist_wtd_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x8_avg = aom_dist_wtd_sad8x8_avg_sse2;
+    aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3;
+    aom_fdct4x4 = aom_fdct4x4_c;
+    if (flags & HAS_SSE2) aom_fdct4x4 = aom_fdct4x4_sse2;
+    aom_fdct4x4_lp = aom_fdct4x4_lp_c;
+    if (flags & HAS_SSE2) aom_fdct4x4_lp = aom_fdct4x4_lp_sse2;
+    aom_fft16x16_float = aom_fft16x16_float_c;
+    if (flags & HAS_SSE2) aom_fft16x16_float = aom_fft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_fft16x16_float = aom_fft16x16_float_avx2;
+    aom_fft32x32_float = aom_fft32x32_float_c;
+    if (flags & HAS_SSE2) aom_fft32x32_float = aom_fft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_fft32x32_float = aom_fft32x32_float_avx2;
+    aom_fft4x4_float = aom_fft4x4_float_c;
+    if (flags & HAS_SSE2) aom_fft4x4_float = aom_fft4x4_float_sse2;
+    aom_fft8x8_float = aom_fft8x8_float_c;
+    if (flags & HAS_SSE2) aom_fft8x8_float = aom_fft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_fft8x8_float = aom_fft8x8_float_avx2;
+    aom_get_blk_sse_sum = aom_get_blk_sse_sum_c;
+    if (flags & HAS_SSE2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_sse2;
+    if (flags & HAS_AVX2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_avx2;
+    aom_get_mb_ss = aom_get_mb_ss_c;
+    if (flags & HAS_SSE2) aom_get_mb_ss = aom_get_mb_ss_sse2;
+    aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_c;
+    if (flags & HAS_SSE2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_avx2;
+    aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_c;
+    if (flags & HAS_SSE2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_avx2;
+    aom_h_predictor_16x16 = aom_h_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x16 = aom_h_predictor_16x16_sse2;
+    aom_h_predictor_16x32 = aom_h_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x32 = aom_h_predictor_16x32_sse2;
+    aom_h_predictor_16x4 = aom_h_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x4 = aom_h_predictor_16x4_sse2;
+    aom_h_predictor_16x64 = aom_h_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x64 = aom_h_predictor_16x64_sse2;
+    aom_h_predictor_16x8 = aom_h_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x8 = aom_h_predictor_16x8_sse2;
+    aom_h_predictor_32x16 = aom_h_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x16 = aom_h_predictor_32x16_sse2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_avx2;
+    aom_h_predictor_32x64 = aom_h_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x64 = aom_h_predictor_32x64_sse2;
+    aom_h_predictor_32x8 = aom_h_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x8 = aom_h_predictor_32x8_sse2;
+    aom_h_predictor_4x16 = aom_h_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x16 = aom_h_predictor_4x16_sse2;
+    aom_h_predictor_4x4 = aom_h_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x4 = aom_h_predictor_4x4_sse2;
+    aom_h_predictor_4x8 = aom_h_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x8 = aom_h_predictor_4x8_sse2;
+    aom_h_predictor_64x16 = aom_h_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_64x16 = aom_h_predictor_64x16_sse2;
+    aom_h_predictor_64x32 = aom_h_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_64x32 = aom_h_predictor_64x32_sse2;
+    aom_h_predictor_64x64 = aom_h_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_h_predictor_64x64 = aom_h_predictor_64x64_sse2;
+    aom_h_predictor_8x16 = aom_h_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x16 = aom_h_predictor_8x16_sse2;
+    aom_h_predictor_8x32 = aom_h_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x32 = aom_h_predictor_8x32_sse2;
+    aom_h_predictor_8x4 = aom_h_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x4 = aom_h_predictor_8x4_sse2;
+    aom_h_predictor_8x8 = aom_h_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x8 = aom_h_predictor_8x8_sse2;
+    aom_hadamard_16x16 = aom_hadamard_16x16_c;
+    if (flags & HAS_SSE2) aom_hadamard_16x16 = aom_hadamard_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_16x16 = aom_hadamard_16x16_avx2;
+    aom_hadamard_32x32 = aom_hadamard_32x32_c;
+    if (flags & HAS_SSE2) aom_hadamard_32x32 = aom_hadamard_32x32_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_32x32 = aom_hadamard_32x32_avx2;
+    aom_hadamard_4x4 = aom_hadamard_4x4_c;
+    if (flags & HAS_SSE2) aom_hadamard_4x4 = aom_hadamard_4x4_sse2;
+    aom_hadamard_8x8 = aom_hadamard_8x8_c;
+    if (flags & HAS_SSE2) aom_hadamard_8x8 = aom_hadamard_8x8_sse2;
+    aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_c;
+    if (flags & HAS_SSE2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_avx2;
+    aom_hadamard_lp_8x8 = aom_hadamard_lp_8x8_c;
+    if (flags & HAS_SSE2) aom_hadamard_lp_8x8 = aom_hadamard_lp_8x8_sse2;
+    aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_c;
+    if (flags & HAS_SSE2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_avx2;
+    aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_sse2;
+    aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
+    aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_sse4_1;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_sse4_1;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_sse4_1;
+    aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_sse4_1;
+    aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_sse4_1;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_sse4_1;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_sse4_1;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_sse4_1;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_sse4_1;
+    aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_sse4_1;
+    aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_sse4_1;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_sse4_1;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_sse4_1;
+    aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_sse4_1;
+    aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_sse4_1;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_sse4_1;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_sse4_1;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_sse4_1;
+    aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_sse4_1;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_sse4_1;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance16x16 = aom_highbd_10_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x16 = aom_highbd_10_sub_pixel_avg_variance16x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x32 = aom_highbd_10_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x32 = aom_highbd_10_sub_pixel_avg_variance16x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x4 = aom_highbd_10_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x4 = aom_highbd_10_sub_pixel_avg_variance16x4_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x64 = aom_highbd_10_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x64 = aom_highbd_10_sub_pixel_avg_variance16x64_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x8 = aom_highbd_10_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x8 = aom_highbd_10_sub_pixel_avg_variance16x8_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x16 = aom_highbd_10_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x16 = aom_highbd_10_sub_pixel_avg_variance32x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x32 = aom_highbd_10_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x32 = aom_highbd_10_sub_pixel_avg_variance32x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x64 = aom_highbd_10_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x64 = aom_highbd_10_sub_pixel_avg_variance32x64_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x8 = aom_highbd_10_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x8 = aom_highbd_10_sub_pixel_avg_variance32x8_sse2;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance64x16 = aom_highbd_10_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance64x16 = aom_highbd_10_sub_pixel_avg_variance64x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance64x32 = aom_highbd_10_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance64x32 = aom_highbd_10_sub_pixel_avg_variance64x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance64x64 = aom_highbd_10_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance64x64 = aom_highbd_10_sub_pixel_avg_variance64x64_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x16 = aom_highbd_10_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x16 = aom_highbd_10_sub_pixel_avg_variance8x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x32 = aom_highbd_10_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x32 = aom_highbd_10_sub_pixel_avg_variance8x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x4 = aom_highbd_10_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x4 = aom_highbd_10_sub_pixel_avg_variance8x4_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x8 = aom_highbd_10_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x8 = aom_highbd_10_sub_pixel_avg_variance8x8_sse2;
+    aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_avx2;
+    aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_avx2;
+    aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_avx2;
+    aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_avx2;
+    aom_highbd_10_sub_pixel_variance16x4 = aom_highbd_10_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x4 = aom_highbd_10_sub_pixel_variance16x4_sse2;
+    aom_highbd_10_sub_pixel_variance16x64 = aom_highbd_10_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x64 = aom_highbd_10_sub_pixel_variance16x64_sse2;
+    aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_avx2;
+    aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_avx2;
+    aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_avx2;
+    aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_avx2;
+    aom_highbd_10_sub_pixel_variance32x8 = aom_highbd_10_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x8 = aom_highbd_10_sub_pixel_variance32x8_sse2;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_avx2;
+    aom_highbd_10_sub_pixel_variance64x16 = aom_highbd_10_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x16 = aom_highbd_10_sub_pixel_variance64x16_sse2;
+    aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_avx2;
+    aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_avx2;
+    aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_avx2;
+    aom_highbd_10_sub_pixel_variance8x32 = aom_highbd_10_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x32 = aom_highbd_10_sub_pixel_variance8x32_sse2;
+    aom_highbd_10_sub_pixel_variance8x4 = aom_highbd_10_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x4 = aom_highbd_10_sub_pixel_variance8x4_sse2;
+    aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_avx2;
+    aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_avx2;
+    aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_avx2;
+    aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_avx2;
+    aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_avx2;
+    aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_avx2;
+    aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_avx2;
+    aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_avx2;
+    aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_avx2;
+    aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_avx2;
+    aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_avx2;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_sse4_1;
+    aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_avx2;
+    aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_avx2;
+    aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_avx2;
+    aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_avx2;
+    aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_avx2;
+    aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_avx2;
+    aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_avx2;
+    aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_12_mse16x16 = aom_highbd_12_mse16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_mse16x16 = aom_highbd_12_mse16x16_sse2;
+    aom_highbd_12_mse8x8 = aom_highbd_12_mse8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_mse8x8 = aom_highbd_12_mse8x8_sse2;
+    aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_sse4_1;
+    aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_sse4_1;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_sse4_1;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_sse4_1;
+    aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_sse4_1;
+    aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_sse4_1;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_sse4_1;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_sse4_1;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_sse4_1;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_sse4_1;
+    aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_sse4_1;
+    aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_sse4_1;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_sse4_1;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_sse4_1;
+    aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_sse4_1;
+    aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_sse4_1;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_sse4_1;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_sse4_1;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_sse4_1;
+    aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_sse4_1;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_sse4_1;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance16x16 = aom_highbd_12_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x16 = aom_highbd_12_sub_pixel_avg_variance16x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x32 = aom_highbd_12_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x32 = aom_highbd_12_sub_pixel_avg_variance16x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x4 = aom_highbd_12_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x4 = aom_highbd_12_sub_pixel_avg_variance16x4_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x64 = aom_highbd_12_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x64 = aom_highbd_12_sub_pixel_avg_variance16x64_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x8 = aom_highbd_12_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x8 = aom_highbd_12_sub_pixel_avg_variance16x8_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x16 = aom_highbd_12_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x16 = aom_highbd_12_sub_pixel_avg_variance32x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x32 = aom_highbd_12_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x32 = aom_highbd_12_sub_pixel_avg_variance32x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x64 = aom_highbd_12_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x64 = aom_highbd_12_sub_pixel_avg_variance32x64_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x8 = aom_highbd_12_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x8 = aom_highbd_12_sub_pixel_avg_variance32x8_sse2;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance64x16 = aom_highbd_12_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance64x16 = aom_highbd_12_sub_pixel_avg_variance64x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance64x32 = aom_highbd_12_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance64x32 = aom_highbd_12_sub_pixel_avg_variance64x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance64x64 = aom_highbd_12_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance64x64 = aom_highbd_12_sub_pixel_avg_variance64x64_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x16 = aom_highbd_12_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x16 = aom_highbd_12_sub_pixel_avg_variance8x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x32 = aom_highbd_12_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x32 = aom_highbd_12_sub_pixel_avg_variance8x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x4 = aom_highbd_12_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x4 = aom_highbd_12_sub_pixel_avg_variance8x4_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x8 = aom_highbd_12_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x8 = aom_highbd_12_sub_pixel_avg_variance8x8_sse2;
+    aom_highbd_12_sub_pixel_variance128x128 = aom_highbd_12_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance128x128 = aom_highbd_12_sub_pixel_variance128x128_sse2;
+    aom_highbd_12_sub_pixel_variance128x64 = aom_highbd_12_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance128x64 = aom_highbd_12_sub_pixel_variance128x64_sse2;
+    aom_highbd_12_sub_pixel_variance16x16 = aom_highbd_12_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x16 = aom_highbd_12_sub_pixel_variance16x16_sse2;
+    aom_highbd_12_sub_pixel_variance16x32 = aom_highbd_12_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x32 = aom_highbd_12_sub_pixel_variance16x32_sse2;
+    aom_highbd_12_sub_pixel_variance16x4 = aom_highbd_12_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x4 = aom_highbd_12_sub_pixel_variance16x4_sse2;
+    aom_highbd_12_sub_pixel_variance16x64 = aom_highbd_12_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x64 = aom_highbd_12_sub_pixel_variance16x64_sse2;
+    aom_highbd_12_sub_pixel_variance16x8 = aom_highbd_12_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x8 = aom_highbd_12_sub_pixel_variance16x8_sse2;
+    aom_highbd_12_sub_pixel_variance32x16 = aom_highbd_12_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x16 = aom_highbd_12_sub_pixel_variance32x16_sse2;
+    aom_highbd_12_sub_pixel_variance32x32 = aom_highbd_12_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x32 = aom_highbd_12_sub_pixel_variance32x32_sse2;
+    aom_highbd_12_sub_pixel_variance32x64 = aom_highbd_12_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x64 = aom_highbd_12_sub_pixel_variance32x64_sse2;
+    aom_highbd_12_sub_pixel_variance32x8 = aom_highbd_12_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x8 = aom_highbd_12_sub_pixel_variance32x8_sse2;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_variance64x128 = aom_highbd_12_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x128 = aom_highbd_12_sub_pixel_variance64x128_sse2;
+    aom_highbd_12_sub_pixel_variance64x16 = aom_highbd_12_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x16 = aom_highbd_12_sub_pixel_variance64x16_sse2;
+    aom_highbd_12_sub_pixel_variance64x32 = aom_highbd_12_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x32 = aom_highbd_12_sub_pixel_variance64x32_sse2;
+    aom_highbd_12_sub_pixel_variance64x64 = aom_highbd_12_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x64 = aom_highbd_12_sub_pixel_variance64x64_sse2;
+    aom_highbd_12_sub_pixel_variance8x16 = aom_highbd_12_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x16 = aom_highbd_12_sub_pixel_variance8x16_sse2;
+    aom_highbd_12_sub_pixel_variance8x32 = aom_highbd_12_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x32 = aom_highbd_12_sub_pixel_variance8x32_sse2;
+    aom_highbd_12_sub_pixel_variance8x4 = aom_highbd_12_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x4 = aom_highbd_12_sub_pixel_variance8x4_sse2;
+    aom_highbd_12_sub_pixel_variance8x8 = aom_highbd_12_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x8 = aom_highbd_12_sub_pixel_variance8x8_sse2;
+    aom_highbd_12_variance128x128 = aom_highbd_12_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance128x128 = aom_highbd_12_variance128x128_sse2;
+    aom_highbd_12_variance128x64 = aom_highbd_12_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance128x64 = aom_highbd_12_variance128x64_sse2;
+    aom_highbd_12_variance16x16 = aom_highbd_12_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x16 = aom_highbd_12_variance16x16_sse2;
+    aom_highbd_12_variance16x32 = aom_highbd_12_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x32 = aom_highbd_12_variance16x32_sse2;
+    aom_highbd_12_variance16x64 = aom_highbd_12_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x64 = aom_highbd_12_variance16x64_sse2;
+    aom_highbd_12_variance16x8 = aom_highbd_12_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x8 = aom_highbd_12_variance16x8_sse2;
+    aom_highbd_12_variance32x16 = aom_highbd_12_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x16 = aom_highbd_12_variance32x16_sse2;
+    aom_highbd_12_variance32x32 = aom_highbd_12_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x32 = aom_highbd_12_variance32x32_sse2;
+    aom_highbd_12_variance32x64 = aom_highbd_12_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x64 = aom_highbd_12_variance32x64_sse2;
+    aom_highbd_12_variance32x8 = aom_highbd_12_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x8 = aom_highbd_12_variance32x8_sse2;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_sse4_1;
+    aom_highbd_12_variance64x128 = aom_highbd_12_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x128 = aom_highbd_12_variance64x128_sse2;
+    aom_highbd_12_variance64x16 = aom_highbd_12_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x16 = aom_highbd_12_variance64x16_sse2;
+    aom_highbd_12_variance64x32 = aom_highbd_12_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x32 = aom_highbd_12_variance64x32_sse2;
+    aom_highbd_12_variance64x64 = aom_highbd_12_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x64 = aom_highbd_12_variance64x64_sse2;
+    aom_highbd_12_variance8x16 = aom_highbd_12_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance8x16 = aom_highbd_12_variance8x16_sse2;
+    aom_highbd_12_variance8x32 = aom_highbd_12_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance8x32 = aom_highbd_12_variance8x32_sse2;
+    aom_highbd_12_variance8x8 = aom_highbd_12_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance8x8 = aom_highbd_12_variance8x8_sse2;
+    aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_8_mse16x16 = aom_highbd_8_mse16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_mse16x16 = aom_highbd_8_mse16x16_sse2;
+    aom_highbd_8_mse8x8 = aom_highbd_8_mse8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_mse8x8 = aom_highbd_8_mse8x8_sse2;
+    aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_sse4_1;
+    aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_sse4_1;
+    aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_sse4_1;
+    aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_sse4_1;
+    aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_sse4_1;
+    aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_sse4_1;
+    aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_sse4_1;
+    aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_sse4_1;
+    aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_sse4_1;
+    aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_sse4_1;
+    aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_sse4_1;
+    aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_sse4_1;
+    aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_sse4_1;
+    aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_sse4_1;
+    aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_sse4_1;
+    aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_sse4_1;
+    aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_sse4_1;
+    aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_sse4_1;
+    aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_sse4_1;
+    aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_sse4_1;
+    aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_sse4_1;
+    aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance16x16 = aom_highbd_8_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x16 = aom_highbd_8_sub_pixel_avg_variance16x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x32 = aom_highbd_8_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x32 = aom_highbd_8_sub_pixel_avg_variance16x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x4 = aom_highbd_8_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x4 = aom_highbd_8_sub_pixel_avg_variance16x4_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x64 = aom_highbd_8_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x64 = aom_highbd_8_sub_pixel_avg_variance16x64_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x8 = aom_highbd_8_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x8 = aom_highbd_8_sub_pixel_avg_variance16x8_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x16 = aom_highbd_8_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x16 = aom_highbd_8_sub_pixel_avg_variance32x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x32 = aom_highbd_8_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x32 = aom_highbd_8_sub_pixel_avg_variance32x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x64 = aom_highbd_8_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x64 = aom_highbd_8_sub_pixel_avg_variance32x64_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x8 = aom_highbd_8_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x8 = aom_highbd_8_sub_pixel_avg_variance32x8_sse2;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance64x16 = aom_highbd_8_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance64x16 = aom_highbd_8_sub_pixel_avg_variance64x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance64x32 = aom_highbd_8_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance64x32 = aom_highbd_8_sub_pixel_avg_variance64x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance64x64 = aom_highbd_8_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance64x64 = aom_highbd_8_sub_pixel_avg_variance64x64_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x16 = aom_highbd_8_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x16 = aom_highbd_8_sub_pixel_avg_variance8x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x32 = aom_highbd_8_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x32 = aom_highbd_8_sub_pixel_avg_variance8x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x4 = aom_highbd_8_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x4 = aom_highbd_8_sub_pixel_avg_variance8x4_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x8 = aom_highbd_8_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x8 = aom_highbd_8_sub_pixel_avg_variance8x8_sse2;
+    aom_highbd_8_sub_pixel_variance128x128 = aom_highbd_8_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance128x128 = aom_highbd_8_sub_pixel_variance128x128_sse2;
+    aom_highbd_8_sub_pixel_variance128x64 = aom_highbd_8_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance128x64 = aom_highbd_8_sub_pixel_variance128x64_sse2;
+    aom_highbd_8_sub_pixel_variance16x16 = aom_highbd_8_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x16 = aom_highbd_8_sub_pixel_variance16x16_sse2;
+    aom_highbd_8_sub_pixel_variance16x32 = aom_highbd_8_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x32 = aom_highbd_8_sub_pixel_variance16x32_sse2;
+    aom_highbd_8_sub_pixel_variance16x4 = aom_highbd_8_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x4 = aom_highbd_8_sub_pixel_variance16x4_sse2;
+    aom_highbd_8_sub_pixel_variance16x64 = aom_highbd_8_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x64 = aom_highbd_8_sub_pixel_variance16x64_sse2;
+    aom_highbd_8_sub_pixel_variance16x8 = aom_highbd_8_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x8 = aom_highbd_8_sub_pixel_variance16x8_sse2;
+    aom_highbd_8_sub_pixel_variance32x16 = aom_highbd_8_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x16 = aom_highbd_8_sub_pixel_variance32x16_sse2;
+    aom_highbd_8_sub_pixel_variance32x32 = aom_highbd_8_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x32 = aom_highbd_8_sub_pixel_variance32x32_sse2;
+    aom_highbd_8_sub_pixel_variance32x64 = aom_highbd_8_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x64 = aom_highbd_8_sub_pixel_variance32x64_sse2;
+    aom_highbd_8_sub_pixel_variance32x8 = aom_highbd_8_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x8 = aom_highbd_8_sub_pixel_variance32x8_sse2;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_variance64x128 = aom_highbd_8_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x128 = aom_highbd_8_sub_pixel_variance64x128_sse2;
+    aom_highbd_8_sub_pixel_variance64x16 = aom_highbd_8_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x16 = aom_highbd_8_sub_pixel_variance64x16_sse2;
+    aom_highbd_8_sub_pixel_variance64x32 = aom_highbd_8_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x32 = aom_highbd_8_sub_pixel_variance64x32_sse2;
+    aom_highbd_8_sub_pixel_variance64x64 = aom_highbd_8_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x64 = aom_highbd_8_sub_pixel_variance64x64_sse2;
+    aom_highbd_8_sub_pixel_variance8x16 = aom_highbd_8_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x16 = aom_highbd_8_sub_pixel_variance8x16_sse2;
+    aom_highbd_8_sub_pixel_variance8x32 = aom_highbd_8_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x32 = aom_highbd_8_sub_pixel_variance8x32_sse2;
+    aom_highbd_8_sub_pixel_variance8x4 = aom_highbd_8_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x4 = aom_highbd_8_sub_pixel_variance8x4_sse2;
+    aom_highbd_8_sub_pixel_variance8x8 = aom_highbd_8_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x8 = aom_highbd_8_sub_pixel_variance8x8_sse2;
+    aom_highbd_8_variance128x128 = aom_highbd_8_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance128x128 = aom_highbd_8_variance128x128_sse2;
+    aom_highbd_8_variance128x64 = aom_highbd_8_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance128x64 = aom_highbd_8_variance128x64_sse2;
+    aom_highbd_8_variance16x16 = aom_highbd_8_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x16 = aom_highbd_8_variance16x16_sse2;
+    aom_highbd_8_variance16x32 = aom_highbd_8_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x32 = aom_highbd_8_variance16x32_sse2;
+    aom_highbd_8_variance16x64 = aom_highbd_8_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x64 = aom_highbd_8_variance16x64_sse2;
+    aom_highbd_8_variance16x8 = aom_highbd_8_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x8 = aom_highbd_8_variance16x8_sse2;
+    aom_highbd_8_variance32x16 = aom_highbd_8_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x16 = aom_highbd_8_variance32x16_sse2;
+    aom_highbd_8_variance32x32 = aom_highbd_8_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x32 = aom_highbd_8_variance32x32_sse2;
+    aom_highbd_8_variance32x64 = aom_highbd_8_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x64 = aom_highbd_8_variance32x64_sse2;
+    aom_highbd_8_variance32x8 = aom_highbd_8_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x8 = aom_highbd_8_variance32x8_sse2;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_sse4_1;
+    aom_highbd_8_variance64x128 = aom_highbd_8_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x128 = aom_highbd_8_variance64x128_sse2;
+    aom_highbd_8_variance64x16 = aom_highbd_8_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x16 = aom_highbd_8_variance64x16_sse2;
+    aom_highbd_8_variance64x32 = aom_highbd_8_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x32 = aom_highbd_8_variance64x32_sse2;
+    aom_highbd_8_variance64x64 = aom_highbd_8_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x64 = aom_highbd_8_variance64x64_sse2;
+    aom_highbd_8_variance8x16 = aom_highbd_8_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance8x16 = aom_highbd_8_variance8x16_sse2;
+    aom_highbd_8_variance8x32 = aom_highbd_8_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance8x32 = aom_highbd_8_variance8x32_sse2;
+    aom_highbd_8_variance8x8 = aom_highbd_8_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance8x8 = aom_highbd_8_variance8x8_sse2;
+    aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_avx2;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_sse4_1;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_sse4_1;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
+    aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_sse2;
+    if (flags & HAS_AVX2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_avx2;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_c;
+    if (flags & HAS_SSE2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_avx2;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_c;
+    if (flags & HAS_SSE2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_avx2;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_c;
+    if (flags & HAS_SSE2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_dc_128_predictor_16x16 = aom_highbd_dc_128_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_16x16 = aom_highbd_dc_128_predictor_16x16_sse2;
+    aom_highbd_dc_128_predictor_16x32 = aom_highbd_dc_128_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_16x32 = aom_highbd_dc_128_predictor_16x32_sse2;
+    aom_highbd_dc_128_predictor_16x8 = aom_highbd_dc_128_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_16x8 = aom_highbd_dc_128_predictor_16x8_sse2;
+    aom_highbd_dc_128_predictor_32x16 = aom_highbd_dc_128_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_32x16 = aom_highbd_dc_128_predictor_32x16_sse2;
+    aom_highbd_dc_128_predictor_32x32 = aom_highbd_dc_128_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_32x32 = aom_highbd_dc_128_predictor_32x32_sse2;
+    aom_highbd_dc_128_predictor_4x4 = aom_highbd_dc_128_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_4x4 = aom_highbd_dc_128_predictor_4x4_sse2;
+    aom_highbd_dc_128_predictor_4x8 = aom_highbd_dc_128_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_4x8 = aom_highbd_dc_128_predictor_4x8_sse2;
+    aom_highbd_dc_128_predictor_8x16 = aom_highbd_dc_128_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_8x16 = aom_highbd_dc_128_predictor_8x16_sse2;
+    aom_highbd_dc_128_predictor_8x4 = aom_highbd_dc_128_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_8x4 = aom_highbd_dc_128_predictor_8x4_sse2;
+    aom_highbd_dc_128_predictor_8x8 = aom_highbd_dc_128_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_8x8 = aom_highbd_dc_128_predictor_8x8_sse2;
+    aom_highbd_dc_left_predictor_16x16 = aom_highbd_dc_left_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_16x16 = aom_highbd_dc_left_predictor_16x16_sse2;
+    aom_highbd_dc_left_predictor_16x32 = aom_highbd_dc_left_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_16x32 = aom_highbd_dc_left_predictor_16x32_sse2;
+    aom_highbd_dc_left_predictor_16x8 = aom_highbd_dc_left_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_16x8 = aom_highbd_dc_left_predictor_16x8_sse2;
+    aom_highbd_dc_left_predictor_32x16 = aom_highbd_dc_left_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_32x16 = aom_highbd_dc_left_predictor_32x16_sse2;
+    aom_highbd_dc_left_predictor_32x32 = aom_highbd_dc_left_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_32x32 = aom_highbd_dc_left_predictor_32x32_sse2;
+    aom_highbd_dc_left_predictor_4x4 = aom_highbd_dc_left_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_4x4 = aom_highbd_dc_left_predictor_4x4_sse2;
+    aom_highbd_dc_left_predictor_4x8 = aom_highbd_dc_left_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_4x8 = aom_highbd_dc_left_predictor_4x8_sse2;
+    aom_highbd_dc_left_predictor_8x16 = aom_highbd_dc_left_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_8x16 = aom_highbd_dc_left_predictor_8x16_sse2;
+    aom_highbd_dc_left_predictor_8x4 = aom_highbd_dc_left_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_8x4 = aom_highbd_dc_left_predictor_8x4_sse2;
+    aom_highbd_dc_left_predictor_8x8 = aom_highbd_dc_left_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_8x8 = aom_highbd_dc_left_predictor_8x8_sse2;
+    aom_highbd_dc_predictor_16x16 = aom_highbd_dc_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_16x16 = aom_highbd_dc_predictor_16x16_sse2;
+    aom_highbd_dc_predictor_16x32 = aom_highbd_dc_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_16x32 = aom_highbd_dc_predictor_16x32_sse2;
+    aom_highbd_dc_predictor_16x8 = aom_highbd_dc_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_16x8 = aom_highbd_dc_predictor_16x8_sse2;
+    aom_highbd_dc_predictor_32x16 = aom_highbd_dc_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_32x16 = aom_highbd_dc_predictor_32x16_sse2;
+    aom_highbd_dc_predictor_32x32 = aom_highbd_dc_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_32x32 = aom_highbd_dc_predictor_32x32_sse2;
+    aom_highbd_dc_predictor_4x4 = aom_highbd_dc_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_4x4 = aom_highbd_dc_predictor_4x4_sse2;
+    aom_highbd_dc_predictor_4x8 = aom_highbd_dc_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_4x8 = aom_highbd_dc_predictor_4x8_sse2;
+    aom_highbd_dc_predictor_8x16 = aom_highbd_dc_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_8x16 = aom_highbd_dc_predictor_8x16_sse2;
+    aom_highbd_dc_predictor_8x4 = aom_highbd_dc_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_8x4 = aom_highbd_dc_predictor_8x4_sse2;
+    aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_sse2;
+    aom_highbd_dc_top_predictor_16x16 = aom_highbd_dc_top_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_16x16 = aom_highbd_dc_top_predictor_16x16_sse2;
+    aom_highbd_dc_top_predictor_16x32 = aom_highbd_dc_top_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_16x32 = aom_highbd_dc_top_predictor_16x32_sse2;
+    aom_highbd_dc_top_predictor_16x8 = aom_highbd_dc_top_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_16x8 = aom_highbd_dc_top_predictor_16x8_sse2;
+    aom_highbd_dc_top_predictor_32x16 = aom_highbd_dc_top_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_32x16 = aom_highbd_dc_top_predictor_32x16_sse2;
+    aom_highbd_dc_top_predictor_32x32 = aom_highbd_dc_top_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_32x32 = aom_highbd_dc_top_predictor_32x32_sse2;
+    aom_highbd_dc_top_predictor_4x4 = aom_highbd_dc_top_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_4x4 = aom_highbd_dc_top_predictor_4x4_sse2;
+    aom_highbd_dc_top_predictor_4x8 = aom_highbd_dc_top_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_4x8 = aom_highbd_dc_top_predictor_4x8_sse2;
+    aom_highbd_dc_top_predictor_8x16 = aom_highbd_dc_top_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_8x16 = aom_highbd_dc_top_predictor_8x16_sse2;
+    aom_highbd_dc_top_predictor_8x4 = aom_highbd_dc_top_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_8x4 = aom_highbd_dc_top_predictor_8x4_sse2;
+    aom_highbd_dc_top_predictor_8x8 = aom_highbd_dc_top_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_8x8 = aom_highbd_dc_top_predictor_8x8_sse2;
+    aom_highbd_dist_wtd_comp_avg_pred = aom_highbd_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_dist_wtd_comp_avg_pred = aom_highbd_dist_wtd_comp_avg_pred_sse2;
+    aom_highbd_h_predictor_16x16 = aom_highbd_h_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_16x16 = aom_highbd_h_predictor_16x16_sse2;
+    aom_highbd_h_predictor_16x32 = aom_highbd_h_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_16x32 = aom_highbd_h_predictor_16x32_sse2;
+    aom_highbd_h_predictor_16x8 = aom_highbd_h_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_16x8 = aom_highbd_h_predictor_16x8_sse2;
+    aom_highbd_h_predictor_32x16 = aom_highbd_h_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_32x16 = aom_highbd_h_predictor_32x16_sse2;
+    aom_highbd_h_predictor_32x32 = aom_highbd_h_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_32x32 = aom_highbd_h_predictor_32x32_sse2;
+    aom_highbd_h_predictor_4x4 = aom_highbd_h_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_4x4 = aom_highbd_h_predictor_4x4_sse2;
+    aom_highbd_h_predictor_4x8 = aom_highbd_h_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_4x8 = aom_highbd_h_predictor_4x8_sse2;
+    aom_highbd_h_predictor_8x16 = aom_highbd_h_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_8x16 = aom_highbd_h_predictor_8x16_sse2;
+    aom_highbd_h_predictor_8x4 = aom_highbd_h_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_8x4 = aom_highbd_h_predictor_8x4_sse2;
+    aom_highbd_h_predictor_8x8 = aom_highbd_h_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_8x8 = aom_highbd_h_predictor_8x8_sse2;
+    aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_avx2;
+    aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_avx2;
+    aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_avx2;
+    aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_sse2;
+    aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_avx2;
+    aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_sse2;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_avx2;
+    aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_sse2;
+    aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_sse2;
+    aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_sse2;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_avx2;
+    aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_sse2;
+    aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_avx2;
+    aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_sse2;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_avx2;
+    aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_sse2;
+    aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_sse2;
+    aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_sse2;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
+    aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_avx2;
+    aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_avx2;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_avx2;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_avx2;
+    aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_avx2;
+    aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_avx2;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_avx2;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_avx2;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_avx2;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_avx2;
+    aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_avx2;
+    aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_avx2;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_avx2;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_avx2;
+    aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_avx2;
+    aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_avx2;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_avx2;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_avx2;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_avx2;
+    aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_avx2;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_avx2;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_avx2;
+    aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_avx2;
+    aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_avx2;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_avx2;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_avx2;
+    aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_avx2;
+    aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_avx2;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_avx2;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_avx2;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_avx2;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_avx2;
+    aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_avx2;
+    aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_avx2;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_avx2;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_avx2;
+    aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_avx2;
+    aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_avx2;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_avx2;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_avx2;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_avx2;
+    aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_avx2;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_avx2;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_avx2;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b = aom_highbd_quantize_b_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b = aom_highbd_quantize_b_avx2;
+    aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_avx2;
+    aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_avx2;
+    aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_avx2;
+    aom_highbd_quantize_b_64x64_adaptive = aom_highbd_quantize_b_64x64_adaptive_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_64x64_adaptive = aom_highbd_quantize_b_64x64_adaptive_sse2;
+    aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_avx2;
+    aom_highbd_sad128x128 = aom_highbd_sad128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128 = aom_highbd_sad128x128_avx2;
+    aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_avx2;
+    aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_avx2;
+    aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_avx2;
+    aom_highbd_sad128x64 = aom_highbd_sad128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64 = aom_highbd_sad128x64_avx2;
+    aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_avx2;
+    aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_avx2;
+    aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_avx2;
+    aom_highbd_sad16x16 = aom_highbd_sad16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x16 = aom_highbd_sad16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16 = aom_highbd_sad16x16_avx2;
+    aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_avx2;
+    aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_avx2;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_avx2;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x32 = aom_highbd_sad16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32 = aom_highbd_sad16x32_avx2;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_avx2;
+    aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_avx2;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_avx2;
+    aom_highbd_sad16x4 = aom_highbd_sad16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x4 = aom_highbd_sad16x4_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4 = aom_highbd_sad16x4_avx2;
+    aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_avx2;
+    aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_avx2;
+    aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_avx2;
+    aom_highbd_sad16x64 = aom_highbd_sad16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x64 = aom_highbd_sad16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64 = aom_highbd_sad16x64_avx2;
+    aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_avx2;
+    aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_avx2;
+    aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_avx2;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x8 = aom_highbd_sad16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8 = aom_highbd_sad16x8_avx2;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_avx2;
+    aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_avx2;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_avx2;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x16 = aom_highbd_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16 = aom_highbd_sad32x16_avx2;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_avx2;
+    aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_avx2;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_avx2;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x32 = aom_highbd_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32 = aom_highbd_sad32x32_avx2;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_avx2;
+    aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_avx2;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_avx2;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x64 = aom_highbd_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64 = aom_highbd_sad32x64_avx2;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_avx2;
+    aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_avx2;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_avx2;
+    aom_highbd_sad32x8 = aom_highbd_sad32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x8 = aom_highbd_sad32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8 = aom_highbd_sad32x8_avx2;
+    aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_avx2;
+    aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_avx2;
+    aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_avx2;
+    aom_highbd_sad4x16 = aom_highbd_sad4x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x16 = aom_highbd_sad4x16_sse2;
+    aom_highbd_sad4x16_avg = aom_highbd_sad4x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x16_avg = aom_highbd_sad4x16_avg_sse2;
+    aom_highbd_sad4x16x4d = aom_highbd_sad4x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x16x4d = aom_highbd_sad4x16x4d_sse2;
+    aom_highbd_sad4x4 = aom_highbd_sad4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x4 = aom_highbd_sad4x4_sse2;
+    aom_highbd_sad4x4_avg = aom_highbd_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x4_avg = aom_highbd_sad4x4_avg_sse2;
+    aom_highbd_sad4x4x4d = aom_highbd_sad4x4x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x4x4d = aom_highbd_sad4x4x4d_sse2;
+    aom_highbd_sad4x8 = aom_highbd_sad4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x8 = aom_highbd_sad4x8_sse2;
+    aom_highbd_sad4x8_avg = aom_highbd_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x8_avg = aom_highbd_sad4x8_avg_sse2;
+    aom_highbd_sad4x8x4d = aom_highbd_sad4x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x8x4d = aom_highbd_sad4x8x4d_sse2;
+    aom_highbd_sad64x128 = aom_highbd_sad64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128 = aom_highbd_sad64x128_avx2;
+    aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_avx2;
+    aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_avx2;
+    aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_avx2;
+    aom_highbd_sad64x16 = aom_highbd_sad64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x16 = aom_highbd_sad64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16 = aom_highbd_sad64x16_avx2;
+    aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_avx2;
+    aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_avx2;
+    aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_avx2;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x32 = aom_highbd_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32 = aom_highbd_sad64x32_avx2;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_avx2;
+    aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_avx2;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_avx2;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x64 = aom_highbd_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64 = aom_highbd_sad64x64_avx2;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_avx2;
+    aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_avx2;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_avx2;
+    aom_highbd_sad8x16 = aom_highbd_sad8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x16 = aom_highbd_sad8x16_sse2;
+    aom_highbd_sad8x16_avg = aom_highbd_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x16_avg = aom_highbd_sad8x16_avg_sse2;
+    aom_highbd_sad8x16x4d = aom_highbd_sad8x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x16x4d = aom_highbd_sad8x16x4d_sse2;
+    aom_highbd_sad8x32 = aom_highbd_sad8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x32 = aom_highbd_sad8x32_sse2;
+    aom_highbd_sad8x32_avg = aom_highbd_sad8x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x32_avg = aom_highbd_sad8x32_avg_sse2;
+    aom_highbd_sad8x32x4d = aom_highbd_sad8x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x32x4d = aom_highbd_sad8x32x4d_sse2;
+    aom_highbd_sad8x4 = aom_highbd_sad8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x4 = aom_highbd_sad8x4_sse2;
+    aom_highbd_sad8x4_avg = aom_highbd_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x4_avg = aom_highbd_sad8x4_avg_sse2;
+    aom_highbd_sad8x4x4d = aom_highbd_sad8x4x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x4x4d = aom_highbd_sad8x4x4d_sse2;
+    aom_highbd_sad8x8 = aom_highbd_sad8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x8 = aom_highbd_sad8x8_sse2;
+    aom_highbd_sad8x8_avg = aom_highbd_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x8_avg = aom_highbd_sad8x8_avg_sse2;
+    aom_highbd_sad8x8x4d = aom_highbd_sad8x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x8x4d = aom_highbd_sad8x8x4d_sse2;
+    aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_avx2;
+    aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_avx2;
+    aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_avx2;
+    aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_avx2;
+    aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_avx2;
+    aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_avx2;
+    aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_avx2;
+    aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_avx2;
+    aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_avx2;
+    aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_avx2;
+    aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_avx2;
+    aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_avx2;
+    aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_avx2;
+    aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_avx2;
+    aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_avx2;
+    aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_avx2;
+    aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_avx2;
+    aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_avx2;
+    aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_avx2;
+    aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_avx2;
+    aom_highbd_sad_skip_4x16 = aom_highbd_sad_skip_4x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x16 = aom_highbd_sad_skip_4x16_sse2;
+    aom_highbd_sad_skip_4x16x4d = aom_highbd_sad_skip_4x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x16x4d = aom_highbd_sad_skip_4x16x4d_sse2;
+    aom_highbd_sad_skip_4x8 = aom_highbd_sad_skip_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x8 = aom_highbd_sad_skip_4x8_sse2;
+    aom_highbd_sad_skip_4x8x4d = aom_highbd_sad_skip_4x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x8x4d = aom_highbd_sad_skip_4x8x4d_sse2;
+    aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_avx2;
+    aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_avx2;
+    aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_avx2;
+    aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_avx2;
+    aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_avx2;
+    aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_avx2;
+    aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_avx2;
+    aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_avx2;
+    aom_highbd_sad_skip_8x16 = aom_highbd_sad_skip_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x16 = aom_highbd_sad_skip_8x16_sse2;
+    aom_highbd_sad_skip_8x16x4d = aom_highbd_sad_skip_8x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x16x4d = aom_highbd_sad_skip_8x16x4d_sse2;
+    aom_highbd_sad_skip_8x32 = aom_highbd_sad_skip_8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x32 = aom_highbd_sad_skip_8x32_sse2;
+    aom_highbd_sad_skip_8x32x4d = aom_highbd_sad_skip_8x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x32x4d = aom_highbd_sad_skip_8x32x4d_sse2;
+    aom_highbd_sad_skip_8x8 = aom_highbd_sad_skip_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x8 = aom_highbd_sad_skip_8x8_sse2;
+    aom_highbd_sad_skip_8x8x4d = aom_highbd_sad_skip_8x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x8x4d = aom_highbd_sad_skip_8x8x4d_sse2;
+    aom_highbd_sse = aom_highbd_sse_c;
+    if (flags & HAS_SSE4_1) aom_highbd_sse = aom_highbd_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_sse = aom_highbd_sse_avx2;
+    aom_highbd_subtract_block = aom_highbd_subtract_block_c;
+    if (flags & HAS_SSE2) aom_highbd_subtract_block = aom_highbd_subtract_block_sse2;
+    aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_sse2;
+    aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_sse2;
+    aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_sse2;
+    aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_sse2;
+    aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_sse2;
+    aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_sse2;
+    aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_sse2;
+    aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_sse2;
+    aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_sse2;
+    aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_sse2;
+    aom_ifft16x16_float = aom_ifft16x16_float_c;
+    if (flags & HAS_SSE2) aom_ifft16x16_float = aom_ifft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft16x16_float = aom_ifft16x16_float_avx2;
+    aom_ifft32x32_float = aom_ifft32x32_float_c;
+    if (flags & HAS_SSE2) aom_ifft32x32_float = aom_ifft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft32x32_float = aom_ifft32x32_float_avx2;
+    aom_ifft4x4_float = aom_ifft4x4_float_c;
+    if (flags & HAS_SSE2) aom_ifft4x4_float = aom_ifft4x4_float_sse2;
+    aom_ifft8x8_float = aom_ifft8x8_float_c;
+    if (flags & HAS_SSE2) aom_ifft8x8_float = aom_ifft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft8x8_float = aom_ifft8x8_float_avx2;
+    aom_int_pro_col = aom_int_pro_col_c;
+    if (flags & HAS_SSE2) aom_int_pro_col = aom_int_pro_col_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_col = aom_int_pro_col_avx2;
+    aom_int_pro_row = aom_int_pro_row_c;
+    if (flags & HAS_SSE2) aom_int_pro_row = aom_int_pro_row_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_row = aom_int_pro_row_avx2;
+    aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
+    aom_lpf_horizontal_14 = aom_lpf_horizontal_14_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_sse2;
+    aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_sse2;
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2;
+    aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_sse2;
+    aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_sse2;
+    aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_sse2;
+    aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_sse2;
+    aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_sse2;
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2;
+    aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_sse2;
+    aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_sse2;
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2;
+    aom_lpf_vertical_14 = aom_lpf_vertical_14_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_14 = aom_lpf_vertical_14_sse2;
+    aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_sse2;
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_avx2;
+    aom_lpf_vertical_4 = aom_lpf_vertical_4_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_4 = aom_lpf_vertical_4_sse2;
+    aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_sse2;
+    aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_sse2;
+    aom_lpf_vertical_6 = aom_lpf_vertical_6_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_6 = aom_lpf_vertical_6_sse2;
+    aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_sse2;
+    aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_sse2;
+    aom_lpf_vertical_8 = aom_lpf_vertical_8_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_8 = aom_lpf_vertical_8_sse2;
+    aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_sse2;
+    aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_sse2;
+    aom_masked_sad128x128 = aom_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128 = aom_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x128 = aom_masked_sad128x128_avx2;
+    aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_ssse3;
+    aom_masked_sad128x64 = aom_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64 = aom_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x64 = aom_masked_sad128x64_avx2;
+    aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_ssse3;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x16 = aom_masked_sad16x16_avx2;
+    aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_ssse3;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32 = aom_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x32 = aom_masked_sad16x32_avx2;
+    aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_ssse3;
+    aom_masked_sad16x4 = aom_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4 = aom_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x4 = aom_masked_sad16x4_avx2;
+    aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_ssse3;
+    aom_masked_sad16x64 = aom_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64 = aom_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x64 = aom_masked_sad16x64_avx2;
+    aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_ssse3;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8 = aom_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x8 = aom_masked_sad16x8_avx2;
+    aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_ssse3;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16 = aom_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x16 = aom_masked_sad32x16_avx2;
+    aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_ssse3;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32 = aom_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x32 = aom_masked_sad32x32_avx2;
+    aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_ssse3;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64 = aom_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x64 = aom_masked_sad32x64_avx2;
+    aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_ssse3;
+    aom_masked_sad32x8 = aom_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8 = aom_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x8 = aom_masked_sad32x8_avx2;
+    aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_ssse3;
+    aom_masked_sad4x16 = aom_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16 = aom_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x16 = aom_masked_sad4x16_avx2;
+    aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_ssse3;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4 = aom_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x4 = aom_masked_sad4x4_avx2;
+    aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_ssse3;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8 = aom_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x8 = aom_masked_sad4x8_avx2;
+    aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_ssse3;
+    aom_masked_sad64x128 = aom_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128 = aom_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x128 = aom_masked_sad64x128_avx2;
+    aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_ssse3;
+    aom_masked_sad64x16 = aom_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16 = aom_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x16 = aom_masked_sad64x16_avx2;
+    aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_ssse3;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32 = aom_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x32 = aom_masked_sad64x32_avx2;
+    aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_ssse3;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64 = aom_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x64 = aom_masked_sad64x64_avx2;
+    aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_ssse3;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16 = aom_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x16 = aom_masked_sad8x16_avx2;
+    aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_ssse3;
+    aom_masked_sad8x32 = aom_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32 = aom_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x32 = aom_masked_sad8x32_avx2;
+    aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_ssse3;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4 = aom_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x4 = aom_masked_sad8x4_avx2;
+    aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_ssse3;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8 = aom_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x8 = aom_masked_sad8x8_avx2;
+    aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_ssse3;
+    aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_ssse3;
+    aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_ssse3;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_ssse3;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_ssse3;
+    aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_ssse3;
+    aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_ssse3;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_ssse3;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_ssse3;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_ssse3;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_ssse3;
+    aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_ssse3;
+    aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_ssse3;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_ssse3;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_ssse3;
+    aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_ssse3;
+    aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_ssse3;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_ssse3;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_ssse3;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_ssse3;
+    aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_ssse3;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_ssse3;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_ssse3;
+    aom_minmax_8x8 = aom_minmax_8x8_c;
+    if (flags & HAS_SSE2) aom_minmax_8x8 = aom_minmax_8x8_sse2;
+    aom_mse16x16 = aom_mse16x16_c;
+    if (flags & HAS_SSE2) aom_mse16x16 = aom_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
+    aom_mse16x8 = aom_mse16x8_c;
+    if (flags & HAS_SSE2) aom_mse16x8 = aom_mse16x8_sse2;
+    aom_mse8x16 = aom_mse8x16_c;
+    if (flags & HAS_SSE2) aom_mse8x16 = aom_mse8x16_sse2;
+    aom_mse8x8 = aom_mse8x8_c;
+    if (flags & HAS_SSE2) aom_mse8x8 = aom_mse8x8_sse2;
+    aom_mse_16xh_16bit = aom_mse_16xh_16bit_c;
+    if (flags & HAS_SSE2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_avx2;
+    aom_mse_wxh_16bit = aom_mse_wxh_16bit_c;
+    if (flags & HAS_SSE2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_avx2;
+    aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_c;
+    if (flags & HAS_SSE2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_avx2;
+    aom_obmc_sad128x128 = aom_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x128 = aom_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x128 = aom_obmc_sad128x128_avx2;
+    aom_obmc_sad128x64 = aom_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x64 = aom_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x64 = aom_obmc_sad128x64_avx2;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x16 = aom_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x16 = aom_obmc_sad16x16_avx2;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x32 = aom_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x32 = aom_obmc_sad16x32_avx2;
+    aom_obmc_sad16x4 = aom_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x4 = aom_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x4 = aom_obmc_sad16x4_avx2;
+    aom_obmc_sad16x64 = aom_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x64 = aom_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x64 = aom_obmc_sad16x64_avx2;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x8 = aom_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x8 = aom_obmc_sad16x8_avx2;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x16 = aom_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x16 = aom_obmc_sad32x16_avx2;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x32 = aom_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x32 = aom_obmc_sad32x32_avx2;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x64 = aom_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x64 = aom_obmc_sad32x64_avx2;
+    aom_obmc_sad32x8 = aom_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x8 = aom_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x8 = aom_obmc_sad32x8_avx2;
+    aom_obmc_sad4x16 = aom_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x16 = aom_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x16 = aom_obmc_sad4x16_avx2;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x4 = aom_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x4 = aom_obmc_sad4x4_avx2;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x8 = aom_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x8 = aom_obmc_sad4x8_avx2;
+    aom_obmc_sad64x128 = aom_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x128 = aom_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x128 = aom_obmc_sad64x128_avx2;
+    aom_obmc_sad64x16 = aom_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x16 = aom_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x16 = aom_obmc_sad64x16_avx2;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x32 = aom_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x32 = aom_obmc_sad64x32_avx2;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x64 = aom_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x64 = aom_obmc_sad64x64_avx2;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x16 = aom_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x16 = aom_obmc_sad8x16_avx2;
+    aom_obmc_sad8x32 = aom_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x32 = aom_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x32 = aom_obmc_sad8x32_avx2;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x4 = aom_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x4 = aom_obmc_sad8x4_avx2;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x8 = aom_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x8 = aom_obmc_sad8x8_avx2;
+    aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_sse4_1;
+    aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_sse4_1;
+    aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_sse4_1;
+    aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_sse4_1;
+    aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_sse4_1;
+    aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_sse4_1;
+    aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_sse4_1;
+    aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_sse4_1;
+    aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_sse4_1;
+    aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_sse4_1;
+    aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_sse4_1;
+    aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_sse4_1;
+    aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_sse4_1;
+    aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_sse4_1;
+    aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_sse4_1;
+    aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_sse4_1;
+    aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_sse4_1;
+    aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_sse4_1;
+    aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_sse4_1;
+    aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_sse4_1;
+    aom_obmc_variance128x128 = aom_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x128 = aom_obmc_variance128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x128 = aom_obmc_variance128x128_avx2;
+    aom_obmc_variance128x64 = aom_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x64 = aom_obmc_variance128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x64 = aom_obmc_variance128x64_avx2;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x16 = aom_obmc_variance16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x16 = aom_obmc_variance16x16_avx2;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x32 = aom_obmc_variance16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x32 = aom_obmc_variance16x32_avx2;
+    aom_obmc_variance16x4 = aom_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x4 = aom_obmc_variance16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x4 = aom_obmc_variance16x4_avx2;
+    aom_obmc_variance16x64 = aom_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x64 = aom_obmc_variance16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x64 = aom_obmc_variance16x64_avx2;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x8 = aom_obmc_variance16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x8 = aom_obmc_variance16x8_avx2;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x16 = aom_obmc_variance32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x16 = aom_obmc_variance32x16_avx2;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x32 = aom_obmc_variance32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x32 = aom_obmc_variance32x32_avx2;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x64 = aom_obmc_variance32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x64 = aom_obmc_variance32x64_avx2;
+    aom_obmc_variance32x8 = aom_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x8 = aom_obmc_variance32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x8 = aom_obmc_variance32x8_avx2;
+    aom_obmc_variance4x16 = aom_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x16 = aom_obmc_variance4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x16 = aom_obmc_variance4x16_avx2;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x4 = aom_obmc_variance4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x4 = aom_obmc_variance4x4_avx2;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x8 = aom_obmc_variance4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x8 = aom_obmc_variance4x8_avx2;
+    aom_obmc_variance64x128 = aom_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x128 = aom_obmc_variance64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x128 = aom_obmc_variance64x128_avx2;
+    aom_obmc_variance64x16 = aom_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x16 = aom_obmc_variance64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x16 = aom_obmc_variance64x16_avx2;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x32 = aom_obmc_variance64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x32 = aom_obmc_variance64x32_avx2;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x64 = aom_obmc_variance64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x64 = aom_obmc_variance64x64_avx2;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x16 = aom_obmc_variance8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x16 = aom_obmc_variance8x16_avx2;
+    aom_obmc_variance8x32 = aom_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x32 = aom_obmc_variance8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x32 = aom_obmc_variance8x32_avx2;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x4 = aom_obmc_variance8x4_avx2;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x8 = aom_obmc_variance8x8_avx2;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_avx2;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_avx2;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_avx2;
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_avx2;
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_avx2;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3;
+    aom_quantize_b = aom_quantize_b_c;
+    if (flags & HAS_SSE2) aom_quantize_b = aom_quantize_b_sse2;
+    if (flags & HAS_AVX) aom_quantize_b = aom_quantize_b_avx;
+    if (flags & HAS_AVX2) aom_quantize_b = aom_quantize_b_avx2;
+    aom_quantize_b_32x32 = aom_quantize_b_32x32_c;
+    if (flags & HAS_AVX) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx;
+    if (flags & HAS_AVX2) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx2;
+    aom_quantize_b_32x32_adaptive = aom_quantize_b_32x32_adaptive_c;
+    if (flags & HAS_SSE2) aom_quantize_b_32x32_adaptive = aom_quantize_b_32x32_adaptive_sse2;
+    aom_quantize_b_64x64 = aom_quantize_b_64x64_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_64x64 = aom_quantize_b_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_quantize_b_64x64 = aom_quantize_b_64x64_avx2;
+    aom_quantize_b_64x64_adaptive = aom_quantize_b_64x64_adaptive_c;
+    if (flags & HAS_SSE2) aom_quantize_b_64x64_adaptive = aom_quantize_b_64x64_adaptive_sse2;
+    aom_quantize_b_adaptive = aom_quantize_b_adaptive_c;
+    if (flags & HAS_SSE2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_avx2;
+    aom_sad128x128 = aom_sad128x128_c;
+    if (flags & HAS_SSE2) aom_sad128x128 = aom_sad128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128 = aom_sad128x128_avx2;
+    aom_sad128x128_avg = aom_sad128x128_avg_c;
+    if (flags & HAS_SSE2) aom_sad128x128_avg = aom_sad128x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128_avg = aom_sad128x128_avg_avx2;
+    aom_sad128x128x3d = aom_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x128x3d = aom_sad128x128x3d_avx2;
+    aom_sad128x128x4d = aom_sad128x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad128x128x4d = aom_sad128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128x4d = aom_sad128x128x4d_avx2;
+    aom_sad128x64 = aom_sad128x64_c;
+    if (flags & HAS_SSE2) aom_sad128x64 = aom_sad128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64 = aom_sad128x64_avx2;
+    aom_sad128x64_avg = aom_sad128x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad128x64_avg = aom_sad128x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64_avg = aom_sad128x64_avg_avx2;
+    aom_sad128x64x3d = aom_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x64x3d = aom_sad128x64x3d_avx2;
+    aom_sad128x64x4d = aom_sad128x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad128x64x4d = aom_sad128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64x4d = aom_sad128x64x4d_avx2;
+    aom_sad16x16 = aom_sad16x16_c;
+    if (flags & HAS_SSE2) aom_sad16x16 = aom_sad16x16_sse2;
+    aom_sad16x16_avg = aom_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x16_avg = aom_sad16x16_avg_sse2;
+    aom_sad16x16x3d = aom_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x16x3d = aom_sad16x16x3d_avx2;
+    aom_sad16x16x4d = aom_sad16x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x16x4d = aom_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x16x4d = aom_sad16x16x4d_avx2;
+    aom_sad16x32 = aom_sad16x32_c;
+    if (flags & HAS_SSE2) aom_sad16x32 = aom_sad16x32_sse2;
+    aom_sad16x32_avg = aom_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x32_avg = aom_sad16x32_avg_sse2;
+    aom_sad16x32x3d = aom_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
+    aom_sad16x32x4d = aom_sad16x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x32x4d = aom_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4 = aom_sad16x4_c;
+    if (flags & HAS_SSE2) aom_sad16x4 = aom_sad16x4_sse2;
+    aom_sad16x4_avg = aom_sad16x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x4_avg = aom_sad16x4_avg_sse2;
+    aom_sad16x4x4d = aom_sad16x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x4x4d = aom_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
+    aom_sad16x64 = aom_sad16x64_c;
+    if (flags & HAS_SSE2) aom_sad16x64 = aom_sad16x64_sse2;
+    aom_sad16x64_avg = aom_sad16x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x64_avg = aom_sad16x64_avg_sse2;
+    aom_sad16x64x3d = aom_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x64x3d = aom_sad16x64x3d_avx2;
+    aom_sad16x64x4d = aom_sad16x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x64x4d = aom_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x64x4d = aom_sad16x64x4d_avx2;
+    aom_sad16x8 = aom_sad16x8_c;
+    if (flags & HAS_SSE2) aom_sad16x8 = aom_sad16x8_sse2;
+    aom_sad16x8_avg = aom_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x8_avg = aom_sad16x8_avg_sse2;
+    aom_sad16x8x3d = aom_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x8x3d = aom_sad16x8x3d_avx2;
+    aom_sad16x8x4d = aom_sad16x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x8x4d = aom_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x8x4d = aom_sad16x8x4d_avx2;
+    aom_sad32x16 = aom_sad32x16_c;
+    if (flags & HAS_SSE2) aom_sad32x16 = aom_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16 = aom_sad32x16_avx2;
+    aom_sad32x16_avg = aom_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x16_avg = aom_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16_avg = aom_sad32x16_avg_avx2;
+    aom_sad32x16x3d = aom_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x16x3d = aom_sad32x16x3d_avx2;
+    aom_sad32x16x4d = aom_sad32x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x16x4d = aom_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16x4d = aom_sad32x16x4d_avx2;
+    aom_sad32x32 = aom_sad32x32_c;
+    if (flags & HAS_SSE2) aom_sad32x32 = aom_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32 = aom_sad32x32_avx2;
+    aom_sad32x32_avg = aom_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x32_avg = aom_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32_avg = aom_sad32x32_avg_avx2;
+    aom_sad32x32x3d = aom_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x32x3d = aom_sad32x32x3d_avx2;
+    aom_sad32x32x4d = aom_sad32x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x32x4d = aom_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32x4d = aom_sad32x32x4d_avx2;
+    aom_sad32x64 = aom_sad32x64_c;
+    if (flags & HAS_SSE2) aom_sad32x64 = aom_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64 = aom_sad32x64_avx2;
+    aom_sad32x64_avg = aom_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x64_avg = aom_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64_avg = aom_sad32x64_avg_avx2;
+    aom_sad32x64x3d = aom_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x64x3d = aom_sad32x64x3d_avx2;
+    aom_sad32x64x4d = aom_sad32x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x64x4d = aom_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64x4d = aom_sad32x64x4d_avx2;
+    aom_sad32x8 = aom_sad32x8_c;
+    if (flags & HAS_SSE2) aom_sad32x8 = aom_sad32x8_sse2;
+    aom_sad32x8_avg = aom_sad32x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x8_avg = aom_sad32x8_avg_sse2;
+    aom_sad32x8x3d = aom_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x8x3d = aom_sad32x8x3d_avx2;
+    aom_sad32x8x4d = aom_sad32x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x8x4d = aom_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x8x4d = aom_sad32x8x4d_avx2;
+    aom_sad4x16 = aom_sad4x16_c;
+    if (flags & HAS_SSE2) aom_sad4x16 = aom_sad4x16_sse2;
+    aom_sad4x16_avg = aom_sad4x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x16_avg = aom_sad4x16_avg_sse2;
+    aom_sad4x16x4d = aom_sad4x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x16x4d = aom_sad4x16x4d_sse2;
+    aom_sad4x4 = aom_sad4x4_c;
+    if (flags & HAS_SSE2) aom_sad4x4 = aom_sad4x4_sse2;
+    aom_sad4x4_avg = aom_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x4_avg = aom_sad4x4_avg_sse2;
+    aom_sad4x4x4d = aom_sad4x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x4x4d = aom_sad4x4x4d_sse2;
+    aom_sad4x8 = aom_sad4x8_c;
+    if (flags & HAS_SSE2) aom_sad4x8 = aom_sad4x8_sse2;
+    aom_sad4x8_avg = aom_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x8_avg = aom_sad4x8_avg_sse2;
+    aom_sad4x8x4d = aom_sad4x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x8x4d = aom_sad4x8x4d_sse2;
+    aom_sad64x128 = aom_sad64x128_c;
+    if (flags & HAS_SSE2) aom_sad64x128 = aom_sad64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128 = aom_sad64x128_avx2;
+    aom_sad64x128_avg = aom_sad64x128_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x128_avg = aom_sad64x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128_avg = aom_sad64x128_avg_avx2;
+    aom_sad64x128x3d = aom_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x128x3d = aom_sad64x128x3d_avx2;
+    aom_sad64x128x4d = aom_sad64x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x128x4d = aom_sad64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128x4d = aom_sad64x128x4d_avx2;
+    aom_sad64x16 = aom_sad64x16_c;
+    if (flags & HAS_SSE2) aom_sad64x16 = aom_sad64x16_sse2;
+    aom_sad64x16_avg = aom_sad64x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x16_avg = aom_sad64x16_avg_sse2;
+    aom_sad64x16x3d = aom_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x16x3d = aom_sad64x16x3d_avx2;
+    aom_sad64x16x4d = aom_sad64x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x16x4d = aom_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x16x4d = aom_sad64x16x4d_avx2;
+    aom_sad64x32 = aom_sad64x32_c;
+    if (flags & HAS_SSE2) aom_sad64x32 = aom_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32 = aom_sad64x32_avx2;
+    aom_sad64x32_avg = aom_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x32_avg = aom_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32_avg = aom_sad64x32_avg_avx2;
+    aom_sad64x32x3d = aom_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x32x3d = aom_sad64x32x3d_avx2;
+    aom_sad64x32x4d = aom_sad64x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x32x4d = aom_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32x4d = aom_sad64x32x4d_avx2;
+    aom_sad64x64 = aom_sad64x64_c;
+    if (flags & HAS_SSE2) aom_sad64x64 = aom_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64 = aom_sad64x64_avx2;
+    aom_sad64x64_avg = aom_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x64_avg = aom_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64_avg = aom_sad64x64_avg_avx2;
+    aom_sad64x64x3d = aom_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x64x3d = aom_sad64x64x3d_avx2;
+    aom_sad64x64x4d = aom_sad64x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x64x4d = aom_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64x4d = aom_sad64x64x4d_avx2;
+    aom_sad8x16 = aom_sad8x16_c;
+    if (flags & HAS_SSE2) aom_sad8x16 = aom_sad8x16_sse2;
+    aom_sad8x16_avg = aom_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x16_avg = aom_sad8x16_avg_sse2;
+    aom_sad8x16x4d = aom_sad8x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x16x4d = aom_sad8x16x4d_sse2;
+    aom_sad8x32 = aom_sad8x32_c;
+    if (flags & HAS_SSE2) aom_sad8x32 = aom_sad8x32_sse2;
+    aom_sad8x32_avg = aom_sad8x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x32_avg = aom_sad8x32_avg_sse2;
+    aom_sad8x32x4d = aom_sad8x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x32x4d = aom_sad8x32x4d_sse2;
+    aom_sad8x4 = aom_sad8x4_c;
+    if (flags & HAS_SSE2) aom_sad8x4 = aom_sad8x4_sse2;
+    aom_sad8x4_avg = aom_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x4_avg = aom_sad8x4_avg_sse2;
+    aom_sad8x4x4d = aom_sad8x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x4x4d = aom_sad8x4x4d_sse2;
+    aom_sad8x8 = aom_sad8x8_c;
+    if (flags & HAS_SSE2) aom_sad8x8 = aom_sad8x8_sse2;
+    aom_sad8x8_avg = aom_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x8_avg = aom_sad8x8_avg_sse2;
+    aom_sad8x8x4d = aom_sad8x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x8x4d = aom_sad8x8x4d_sse2;
+    aom_sad_skip_128x128 = aom_sad_skip_128x128_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x128 = aom_sad_skip_128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128 = aom_sad_skip_128x128_avx2;
+    aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_avx2;
+    aom_sad_skip_128x64 = aom_sad_skip_128x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x64 = aom_sad_skip_128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64 = aom_sad_skip_128x64_avx2;
+    aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_avx2;
+    aom_sad_skip_16x16 = aom_sad_skip_16x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x16 = aom_sad_skip_16x16_sse2;
+    aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
+    aom_sad_skip_16x32 = aom_sad_skip_16x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x32 = aom_sad_skip_16x32_sse2;
+    aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x64 = aom_sad_skip_16x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x64 = aom_sad_skip_16x64_sse2;
+    aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
+    aom_sad_skip_16x8 = aom_sad_skip_16x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x8 = aom_sad_skip_16x8_sse2;
+    aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_avx2;
+    aom_sad_skip_32x16 = aom_sad_skip_32x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x16 = aom_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16 = aom_sad_skip_32x16_avx2;
+    aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_avx2;
+    aom_sad_skip_32x32 = aom_sad_skip_32x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x32 = aom_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32 = aom_sad_skip_32x32_avx2;
+    aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_avx2;
+    aom_sad_skip_32x64 = aom_sad_skip_32x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x64 = aom_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64 = aom_sad_skip_32x64_avx2;
+    aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_avx2;
+    aom_sad_skip_32x8 = aom_sad_skip_32x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x8 = aom_sad_skip_32x8_sse2;
+    aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_avx2;
+    aom_sad_skip_4x16 = aom_sad_skip_4x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x16 = aom_sad_skip_4x16_sse2;
+    aom_sad_skip_4x16x4d = aom_sad_skip_4x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x16x4d = aom_sad_skip_4x16x4d_sse2;
+    aom_sad_skip_4x8 = aom_sad_skip_4x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x8 = aom_sad_skip_4x8_sse2;
+    aom_sad_skip_4x8x4d = aom_sad_skip_4x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x8x4d = aom_sad_skip_4x8x4d_sse2;
+    aom_sad_skip_64x128 = aom_sad_skip_64x128_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x128 = aom_sad_skip_64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128 = aom_sad_skip_64x128_avx2;
+    aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_avx2;
+    aom_sad_skip_64x16 = aom_sad_skip_64x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x16 = aom_sad_skip_64x16_sse2;
+    aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_avx2;
+    aom_sad_skip_64x32 = aom_sad_skip_64x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x32 = aom_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32 = aom_sad_skip_64x32_avx2;
+    aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_avx2;
+    aom_sad_skip_64x64 = aom_sad_skip_64x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x64 = aom_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64 = aom_sad_skip_64x64_avx2;
+    aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_avx2;
+    aom_sad_skip_8x16 = aom_sad_skip_8x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x16 = aom_sad_skip_8x16_sse2;
+    aom_sad_skip_8x16x4d = aom_sad_skip_8x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x16x4d = aom_sad_skip_8x16x4d_sse2;
+    aom_sad_skip_8x32 = aom_sad_skip_8x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x32 = aom_sad_skip_8x32_sse2;
+    aom_sad_skip_8x32x4d = aom_sad_skip_8x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x32x4d = aom_sad_skip_8x32x4d_sse2;
+    aom_sad_skip_8x8 = aom_sad_skip_8x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x8 = aom_sad_skip_8x8_sse2;
+    aom_sad_skip_8x8x4d = aom_sad_skip_8x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x8x4d = aom_sad_skip_8x8x4d_sse2;
+    aom_satd = aom_satd_c;
+    if (flags & HAS_SSE2) aom_satd = aom_satd_sse2;
+    if (flags & HAS_AVX2) aom_satd = aom_satd_avx2;
+    aom_satd_lp = aom_satd_lp_c;
+    if (flags & HAS_SSE2) aom_satd_lp = aom_satd_lp_sse2;
+    if (flags & HAS_AVX2) aom_satd_lp = aom_satd_lp_avx2;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3;
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_ssse3;
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_ssse3;
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_ssse3;
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_ssse3;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_ssse3;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_ssse3;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_ssse3;
+    aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
+    aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
+    aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_ssse3;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_ssse3;
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_ssse3;
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_ssse3;
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_ssse3;
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_ssse3;
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_ssse3;
+    aom_sse = aom_sse_c;
+    if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_subtract_block = aom_subtract_block_c;
+    if (flags & HAS_SSE2) aom_subtract_block = aom_subtract_block_sse2;
+    if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_c;
+    if (flags & HAS_SSE2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_avx2;
+    aom_sum_squares_i16 = aom_sum_squares_i16_c;
+    if (flags & HAS_SSE2) aom_sum_squares_i16 = aom_sum_squares_i16_sse2;
+    aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_c;
+    if (flags & HAS_SSE2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_avx2;
+    aom_v_predictor_16x16 = aom_v_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x16 = aom_v_predictor_16x16_sse2;
+    aom_v_predictor_16x32 = aom_v_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x32 = aom_v_predictor_16x32_sse2;
+    aom_v_predictor_16x4 = aom_v_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x4 = aom_v_predictor_16x4_sse2;
+    aom_v_predictor_16x64 = aom_v_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x64 = aom_v_predictor_16x64_sse2;
+    aom_v_predictor_16x8 = aom_v_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x8 = aom_v_predictor_16x8_sse2;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x16 = aom_v_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x16 = aom_v_predictor_32x16_avx2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x32 = aom_v_predictor_32x32_avx2;
+    aom_v_predictor_32x64 = aom_v_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+    aom_v_predictor_32x8 = aom_v_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x8 = aom_v_predictor_32x8_sse2;
+    aom_v_predictor_4x16 = aom_v_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x16 = aom_v_predictor_4x16_sse2;
+    aom_v_predictor_4x4 = aom_v_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x4 = aom_v_predictor_4x4_sse2;
+    aom_v_predictor_4x8 = aom_v_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x8 = aom_v_predictor_4x8_sse2;
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
+    aom_v_predictor_64x32 = aom_v_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
+    aom_v_predictor_64x64 = aom_v_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_v_predictor_64x64 = aom_v_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x64 = aom_v_predictor_64x64_avx2;
+    aom_v_predictor_8x16 = aom_v_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x16 = aom_v_predictor_8x16_sse2;
+    aom_v_predictor_8x32 = aom_v_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x32 = aom_v_predictor_8x32_sse2;
+    aom_v_predictor_8x4 = aom_v_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x4 = aom_v_predictor_8x4_sse2;
+    aom_v_predictor_8x8 = aom_v_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x8 = aom_v_predictor_8x8_sse2;
+    aom_var_2d_u16 = aom_var_2d_u16_c;
+    if (flags & HAS_SSE2) aom_var_2d_u16 = aom_var_2d_u16_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u16 = aom_var_2d_u16_avx2;
+    aom_var_2d_u8 = aom_var_2d_u8_c;
+    if (flags & HAS_SSE2) aom_var_2d_u8 = aom_var_2d_u8_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u8 = aom_var_2d_u8_avx2;
+    aom_variance128x128 = aom_variance128x128_c;
+    if (flags & HAS_SSE2) aom_variance128x128 = aom_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_variance128x128 = aom_variance128x128_avx2;
+    aom_variance128x64 = aom_variance128x64_c;
+    if (flags & HAS_SSE2) aom_variance128x64 = aom_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_variance128x64 = aom_variance128x64_avx2;
+    aom_variance16x16 = aom_variance16x16_c;
+    if (flags & HAS_SSE2) aom_variance16x16 = aom_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
+    aom_variance16x32 = aom_variance16x32_c;
+    if (flags & HAS_SSE2) aom_variance16x32 = aom_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_variance16x32 = aom_variance16x32_avx2;
+    aom_variance16x4 = aom_variance16x4_c;
+    if (flags & HAS_SSE2) aom_variance16x4 = aom_variance16x4_sse2;
+    if (flags & HAS_AVX2) aom_variance16x4 = aom_variance16x4_avx2;
+    aom_variance16x64 = aom_variance16x64_c;
+    if (flags & HAS_SSE2) aom_variance16x64 = aom_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_variance16x64 = aom_variance16x64_avx2;
+    aom_variance16x8 = aom_variance16x8_c;
+    if (flags & HAS_SSE2) aom_variance16x8 = aom_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_variance16x8 = aom_variance16x8_avx2;
+    aom_variance32x16 = aom_variance32x16_c;
+    if (flags & HAS_SSE2) aom_variance32x16 = aom_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_variance32x16 = aom_variance32x16_avx2;
+    aom_variance32x32 = aom_variance32x32_c;
+    if (flags & HAS_SSE2) aom_variance32x32 = aom_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_variance32x32 = aom_variance32x32_avx2;
+    aom_variance32x64 = aom_variance32x64_c;
+    if (flags & HAS_SSE2) aom_variance32x64 = aom_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_variance32x64 = aom_variance32x64_avx2;
+    aom_variance32x8 = aom_variance32x8_c;
+    if (flags & HAS_SSE2) aom_variance32x8 = aom_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_variance32x8 = aom_variance32x8_avx2;
+    aom_variance4x16 = aom_variance4x16_c;
+    if (flags & HAS_SSE2) aom_variance4x16 = aom_variance4x16_sse2;
+    aom_variance4x4 = aom_variance4x4_c;
+    if (flags & HAS_SSE2) aom_variance4x4 = aom_variance4x4_sse2;
+    aom_variance4x8 = aom_variance4x8_c;
+    if (flags & HAS_SSE2) aom_variance4x8 = aom_variance4x8_sse2;
+    aom_variance64x128 = aom_variance64x128_c;
+    if (flags & HAS_SSE2) aom_variance64x128 = aom_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_variance64x128 = aom_variance64x128_avx2;
+    aom_variance64x16 = aom_variance64x16_c;
+    if (flags & HAS_SSE2) aom_variance64x16 = aom_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_variance64x16 = aom_variance64x16_avx2;
+    aom_variance64x32 = aom_variance64x32_c;
+    if (flags & HAS_SSE2) aom_variance64x32 = aom_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_variance64x32 = aom_variance64x32_avx2;
+    aom_variance64x64 = aom_variance64x64_c;
+    if (flags & HAS_SSE2) aom_variance64x64 = aom_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_variance64x64 = aom_variance64x64_avx2;
+    aom_variance8x16 = aom_variance8x16_c;
+    if (flags & HAS_SSE2) aom_variance8x16 = aom_variance8x16_sse2;
+    aom_variance8x32 = aom_variance8x32_c;
+    if (flags & HAS_SSE2) aom_variance8x32 = aom_variance8x32_sse2;
+    aom_variance8x4 = aom_variance8x4_c;
+    if (flags & HAS_SSE2) aom_variance8x4 = aom_variance8x4_sse2;
+    aom_variance8x8 = aom_variance8x8_c;
+    if (flags & HAS_SSE2) aom_variance8x8 = aom_variance8x8_sse2;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
+    if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
+    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
+    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h b/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..3b70fb47c3
--- /dev/null
+++ b/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h
@@ -0,0 +1,103 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/ia32/config/av1_rtcd.h b/media/libaom/config/linux/ia32/config/av1_rtcd.h
new file mode 100644
index 0000000000..3f404f61c8
--- /dev/null
+++ b/media/libaom/config/linux/ia32/config/av1_rtcd.h
@@ -0,0 +1,1272 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+void aom_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+RTCD_EXTERN void (*aom_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_dist_wtd_comp_avg_upsampled_pred_ssse3(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+#define aom_quantize_b_helper aom_quantize_b_helper_c
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+RTCD_EXTERN void (*aom_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_sse4_1(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_avx2(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_avx2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+RTCD_EXTERN int64_t (*av1_block_error_lp)(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim1)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim2)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params_high_bd)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+void av1_cnn_convolve_no_maxpool_padding_valid_avx2(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+RTCD_EXTERN void (*av1_cnn_convolve_no_maxpool_padding_valid)(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+RTCD_EXTERN void (*av1_compute_stats)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+RTCD_EXTERN void (*av1_compute_stats_highbd)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr_intrabc av1_convolve_2d_sr_intrabc_c
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr_intrabc av1_convolve_x_sr_intrabc_c
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+double av1_estimate_noise_from_single_plane_avx2(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+RTCD_EXTERN double (*av1_estimate_noise_from_single_plane)(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void av1_fwht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*av1_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+uint32_t av1_get_crc32c_value_sse4_2(void *crc_calculator, uint8_t *p, size_t length);
+RTCD_EXTERN uint32_t (*av1_get_crc32c_value)(void *crc_calculator, uint8_t *p, size_t length);
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_sse4_1(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_avx2(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+RTCD_EXTERN void (*av1_get_horver_correlation_full)(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+void av1_get_nz_map_contexts_sse2(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+RTCD_EXTERN void (*av1_get_nz_map_contexts)(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_highbd_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+RTCD_EXTERN int64_t (*av1_highbd_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_c
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_c
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr_intrabc av1_highbd_convolve_2d_sr_intrabc_c
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr_intrabc av1_highbd_convolve_x_sr_intrabc_c
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr_intrabc av1_highbd_convolve_y_sr_intrabc_c
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+#define av1_highbd_estimate_noise_from_single_plane av1_highbd_estimate_noise_from_single_plane_c
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+void av1_highbd_filter_intra_edge_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_highbd_filter_intra_edge)(uint16_t *p, int sz, int strength);
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_highbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_sse4_1(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+RTCD_EXTERN void (*av1_highbd_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+void av1_highbd_upsample_intra_edge_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_highbd_upsample_intra_edge)(uint16_t *p, int sz, int bd);
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_lowbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+void av1_nn_fast_softmax_16_sse3(const float *input_nodes, float *output);
+RTCD_EXTERN void (*av1_nn_fast_softmax_16)(const float *input_nodes, float *output);
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_sse3(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_avx2(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+RTCD_EXTERN void (*av1_nn_predict)(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_avx2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_lp)(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit);
+RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit);
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+RTCD_EXTERN void (*av1_txb_init_levels)(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_avx2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+RTCD_EXTERN void (*av1_wedge_compute_delta_squares)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_avx2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+RTCD_EXTERN int8_t (*av1_wedge_sign_from_residuals)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+RTCD_EXTERN uint64_t (*av1_wedge_sse_from_residuals)(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*cfl_get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*cfl_get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*cfl_get_subtract_average_fn)(TX_SIZE tx_size);
+
+void av1_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_sse2;
+    aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+    aom_highbd_comp_avg_upsampled_pred = aom_highbd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_comp_avg_upsampled_pred = aom_highbd_comp_avg_upsampled_pred_sse2;
+    aom_highbd_dist_wtd_comp_avg_upsampled_pred = aom_highbd_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_dist_wtd_comp_avg_upsampled_pred = aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2;
+    aom_highbd_upsampled_pred = aom_highbd_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_upsampled_pred = aom_highbd_upsampled_pred_sse2;
+    aom_upsampled_pred = aom_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_upsampled_pred = aom_upsampled_pred_sse2;
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
+    av1_apply_temporal_filter = av1_apply_temporal_filter_c;
+    if (flags & HAS_SSE2) av1_apply_temporal_filter = av1_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_apply_temporal_filter = av1_apply_temporal_filter_avx2;
+    av1_block_error = av1_block_error_c;
+    if (flags & HAS_SSE2) av1_block_error = av1_block_error_sse2;
+    if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_block_error_lp = av1_block_error_lp_c;
+    if (flags & HAS_SSE2) av1_block_error_lp = av1_block_error_lp_sse2;
+    if (flags & HAS_AVX2) av1_block_error_lp = av1_block_error_lp_avx2;
+    av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
+    av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
+    av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
+    if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
+    av1_calc_indices_dim1 = av1_calc_indices_dim1_c;
+    if (flags & HAS_SSE2) av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
+    av1_calc_indices_dim2 = av1_calc_indices_dim2_c;
+    if (flags & HAS_SSE2) av1_calc_indices_dim2 = av1_calc_indices_dim2_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim2 = av1_calc_indices_dim2_avx2;
+    av1_calc_proj_params = av1_calc_proj_params_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params = av1_calc_proj_params_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params = av1_calc_proj_params_avx2;
+    av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_avx2;
+    av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_c;
+    if (flags & HAS_AVX2) av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_avx2;
+    av1_compute_stats = av1_compute_stats_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats = av1_compute_stats_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats = av1_compute_stats_avx2;
+    av1_compute_stats_highbd = av1_compute_stats_highbd_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats_highbd = av1_compute_stats_highbd_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats_highbd = av1_compute_stats_highbd_avx2;
+    av1_convolve_2d_scale = av1_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
+    av1_convolve_2d_sr = av1_convolve_2d_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
+    av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
+    av1_convolve_x_sr = av1_convolve_x_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_x_sr = av1_convolve_x_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
+    av1_convolve_y_sr = av1_convolve_y_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_y_sr = av1_convolve_y_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
+    av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_avx2;
+    av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_avx2;
+    av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2;
+    av1_dr_prediction_z1 = av1_dr_prediction_z1_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2;
+    av1_dr_prediction_z2 = av1_dr_prediction_z2_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2;
+    av1_dr_prediction_z3 = av1_dr_prediction_z3_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2;
+    av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_c;
+    if (flags & HAS_AVX2) av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_avx2;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
+    av1_filter_intra_predictor = av1_filter_intra_predictor_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_avx2;
+    av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_sse4_1;
+    av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_sse4_1;
+    av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_sse4_1;
+    av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_avx2;
+    av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_sse4_1;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_avx2;
+    av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_sse4_1;
+    av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_sse4_1;
+    av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_sse4_1;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
+    av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_sse4_1;
+    av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_sse4_1;
+    av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_sse4_1;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_avx2;
+    av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_avx2;
+    av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_sse4_1;
+    av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_sse4_1;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_avx2;
+    av1_fwht4x4 = av1_fwht4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwht4x4 = av1_fwht4x4_sse4_1;
+    av1_get_crc32c_value = av1_get_crc32c_value_c;
+    if (flags & HAS_SSE4_2) av1_get_crc32c_value = av1_get_crc32c_value_sse4_2;
+    av1_get_horver_correlation_full = av1_get_horver_correlation_full_c;
+    if (flags & HAS_SSE4_1) av1_get_horver_correlation_full = av1_get_horver_correlation_full_sse4_1;
+    if (flags & HAS_AVX2) av1_get_horver_correlation_full = av1_get_horver_correlation_full_avx2;
+    av1_get_nz_map_contexts = av1_get_nz_map_contexts_c;
+    if (flags & HAS_SSE2) av1_get_nz_map_contexts = av1_get_nz_map_contexts_sse2;
+    av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_c;
+    if (flags & HAS_SSE2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_avx2;
+    av1_highbd_block_error = av1_highbd_block_error_c;
+    if (flags & HAS_SSE2) av1_highbd_block_error = av1_highbd_block_error_sse2;
+    if (flags & HAS_AVX2) av1_highbd_block_error = av1_highbd_block_error_avx2;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
+    av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
+    av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
+    av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
+    av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_avx2;
+    av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_avx2;
+    av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_avx2;
+    av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_avx2;
+    av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_avx2;
+    av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_avx2;
+    av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_avx2;
+    av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_sse4_1;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_sse4_1;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_sse4_1;
+    av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
+    av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1;
+    av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_avx2;
+    av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
+    if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
+    av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_sse4_1;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2;
+    av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
+    if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
+    av1_inv_txfm_add = av1_inv_txfm_add_c;
+    if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
+    if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_c;
+    if (flags & HAS_SSE2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
+    av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_avx2;
+    av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_c;
+    if (flags & HAS_SSE3) av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_sse3;
+    av1_nn_predict = av1_nn_predict_c;
+    if (flags & HAS_SSE3) av1_nn_predict = av1_nn_predict_sse3;
+    if (flags & HAS_AVX2) av1_nn_predict = av1_nn_predict_avx2;
+    av1_quantize_fp = av1_quantize_fp_c;
+    if (flags & HAS_SSE2) av1_quantize_fp = av1_quantize_fp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
+    av1_quantize_fp_64x64 = av1_quantize_fp_64x64_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_64x64 = av1_quantize_fp_64x64_avx2;
+    av1_quantize_lp = av1_quantize_lp_c;
+    if (flags & HAS_SSE2) av1_quantize_lp = av1_quantize_lp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_lp = av1_quantize_lp_avx2;
+    av1_resize_and_extend_frame = av1_resize_and_extend_frame_c;
+    if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
+    av1_txb_init_levels = av1_txb_init_levels_c;
+    if (flags & HAS_SSE4_1) av1_txb_init_levels = av1_txb_init_levels_sse4_1;
+    if (flags & HAS_AVX2) av1_txb_init_levels = av1_txb_init_levels_avx2;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_warp_affine = av1_warp_affine_avx2;
+    av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_c;
+    if (flags & HAS_SSE2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
+    if (flags & HAS_AVX2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
+    av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_c;
+    if (flags & HAS_SSE2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_avx2;
+    av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_c;
+    if (flags & HAS_SSE2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
+    av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_c;
+    if (flags & HAS_SSE2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
+    if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
+    if (flags & HAS_SSE2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
+    if (flags & HAS_SSE2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
+    cdef_filter_16_0 = cdef_filter_16_0_c;
+    if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
+    cdef_filter_16_1 = cdef_filter_16_1_c;
+    if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
+    cdef_filter_16_2 = cdef_filter_16_2_c;
+    if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
+    cdef_filter_16_3 = cdef_filter_16_3_c;
+    if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
+    cdef_filter_8_0 = cdef_filter_8_0_c;
+    if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
+    cdef_filter_8_1 = cdef_filter_8_1_c;
+    if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
+    cdef_filter_8_2 = cdef_filter_8_2_c;
+    if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
+    cdef_filter_8_3 = cdef_filter_8_3_c;
+    if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
+    cdef_find_dir = cdef_find_dir_c;
+    if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
+    cdef_find_dir_dual = cdef_find_dir_dual_c;
+    if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
+    cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
+    cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
+    cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
+    cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
+    cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
+    cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
+    cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_avx2;
+    cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_avx2;
+    cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_c;
+    if (flags & HAS_SSE2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/x64/config/aom_config.asm b/media/libaom/config/linux/x64/config/aom_config.asm
new file mode 100644
index 0000000000..f793ff3c6d
--- /dev/null
+++ b/media/libaom/config/linux/x64/config/aom_config.asm
@@ -0,0 +1,92 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+AOM_ARCH_AARCH64 equ 0
+AOM_ARCH_ARM equ 0
+AOM_ARCH_PPC equ 0
+AOM_ARCH_X86 equ 0
+AOM_ARCH_X86_64 equ 1
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_HIGHBITDEPTH equ 1
+CONFIG_AV1_TEMPORAL_DENOISING equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITRATE_ACCURACY equ 0
+CONFIG_BITRATE_ACCURACY_BL equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_COMPONENT_TIMING equ 0
+CONFIG_COLLECT_PARTITION_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_CWG_C013 equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_EXCLUDE_SIMD_MISMATCH equ 0
+CONFIG_FPMT_TEST equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBVMAF_PSNR_PEAK equ 1
+CONFIG_LIBYUV equ 0
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NN_V2 equ 0
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OPTICAL_FLOW_API equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_OUTPUT_FRAME_SIZE equ 0
+CONFIG_PARTITION_SEARCH_ORDER equ 0
+CONFIG_PIC equ 0
+CONFIG_RATECTRL_LOG equ 0
+CONFIG_RD_COMMAND equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_RT_ML_PARTITIONING equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SALIENCY_MAP equ 0
+CONFIG_SHARED equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_SPEED_STATS equ 0
+CONFIG_TFLITE equ 0
+CONFIG_THREE_PASS equ 0
+CONFIG_TUNE_BUTTERAUGLI equ 0
+CONFIG_TUNE_VMAF equ 0
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+FORCE_HIGHBITDEPTH_DECODING equ 0
+HAVE_ARM_CRC32 equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_FEXCEPT equ 1
+HAVE_MMX equ 1
+HAVE_NEON equ 0
+HAVE_NEON_DOTPROD equ 0
+HAVE_NEON_I8MM equ 0
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 1
+HAVE_SSE4_1 equ 1
+HAVE_SSE4_2 equ 1
+HAVE_SSSE3 equ 1
+HAVE_SVE equ 0
+HAVE_VSX equ 0
+HAVE_WXWIDGETS equ 0
+STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/linux/x64/config/aom_config.h b/media/libaom/config/linux/x64/config/aom_config.h
new file mode 100644
index 0000000000..670d2ffe56
--- /dev/null
+++ b/media/libaom/config/linux/x64/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 0
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 0
+#define AOM_ARCH_X86_64 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 0
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 1
+#define HAVE_NEON 0
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_SSE4_2 1
+#define HAVE_SSSE3 1
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..8e979cc189
--- /dev/null
+++ b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
@@ -0,0 +1,8036 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+#define aom_avg_4x4 aom_avg_4x4_sse2
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+#define aom_avg_8x8 aom_avg_8x8_sse2
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_avx2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_sse2
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_sse2
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_sse2
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_sse2
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_sse2
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_sse2
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_sse2
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_sse2
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x16 aom_dc_predictor_16x16_sse2
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x4 aom_dc_predictor_4x4_sse2
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_sse2
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x8 aom_dc_predictor_8x8_sse2
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_sse2
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_sse2
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_sse2
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_sse2
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_dist_wtd_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x128_avg aom_dist_wtd_sad128x128_avg_sse2
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x64_avg aom_dist_wtd_sad128x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x16_avg aom_dist_wtd_sad16x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x32_avg aom_dist_wtd_sad16x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x4_avg aom_dist_wtd_sad16x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x64_avg aom_dist_wtd_sad16x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x8_avg aom_dist_wtd_sad16x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x16_avg aom_dist_wtd_sad32x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x32_avg aom_dist_wtd_sad32x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x64_avg aom_dist_wtd_sad32x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x8_avg aom_dist_wtd_sad32x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x16_avg aom_dist_wtd_sad4x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x4_avg aom_dist_wtd_sad4x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x8_avg aom_dist_wtd_sad4x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x128_avg aom_dist_wtd_sad64x128_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x16_avg aom_dist_wtd_sad64x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x32_avg aom_dist_wtd_sad64x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x64_avg aom_dist_wtd_sad64x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x16_avg aom_dist_wtd_sad8x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x32_avg aom_dist_wtd_sad8x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x4_avg aom_dist_wtd_sad8x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x8_avg aom_dist_wtd_sad8x8_avg_sse2
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_sse2
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+void aom_fdct4x4_lp_sse2(const int16_t *input, int16_t *output, int stride);
+#define aom_fdct4x4_lp aom_fdct4x4_lp_sse2
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+void aom_fft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_fft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+void aom_fft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_fft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+void aom_fft4x4_float_sse2(const float *input, float *temp, float *output);
+#define aom_fft4x4_float aom_fft4x4_float_sse2
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+void aom_fft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_fft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_sse2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_avx2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+RTCD_EXTERN void (*aom_get_blk_sse_sum)(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_sse2
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+RTCD_EXTERN void (*aom_get_var_sse_sum_16x16_dual)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+RTCD_EXTERN void (*aom_get_var_sse_sum_8x8_quad)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x16 aom_h_predictor_16x16_sse2
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x4 aom_h_predictor_4x4_sse2
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x32 aom_h_predictor_64x32_sse2
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x64 aom_h_predictor_64x64_sse2
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_sse2
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x8 aom_h_predictor_8x8_sse2
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_4x4 aom_hadamard_4x4_sse2
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_8x8 aom_hadamard_8x8_sse2
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+#define aom_hadamard_lp_8x8 aom_hadamard_lp_8x8_sse2
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8_dual)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_sse2
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x128 aom_highbd_10_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x64 aom_highbd_10_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x64 aom_highbd_10_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x128 aom_highbd_10_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x16 aom_highbd_10_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x128 aom_highbd_10_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x64 aom_highbd_10_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x4 aom_highbd_10_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x64 aom_highbd_10_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x8 aom_highbd_10_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x128 aom_highbd_10_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x16 aom_highbd_10_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x32 aom_highbd_10_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x4 aom_highbd_10_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x64 aom_highbd_10_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x8 aom_highbd_10_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x16 aom_highbd_10_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x32 aom_highbd_10_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_sse2
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_sse2
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x128 aom_highbd_12_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x64 aom_highbd_12_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x64 aom_highbd_12_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x128 aom_highbd_12_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x16 aom_highbd_12_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x128 aom_highbd_12_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x64 aom_highbd_12_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x4 aom_highbd_12_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x64 aom_highbd_12_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x8 aom_highbd_12_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x128 aom_highbd_12_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x16 aom_highbd_12_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x32 aom_highbd_12_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x128 aom_highbd_12_sub_pixel_variance128x128_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x64 aom_highbd_12_sub_pixel_variance128x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x4 aom_highbd_12_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x64 aom_highbd_12_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x8 aom_highbd_12_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x128 aom_highbd_12_sub_pixel_variance64x128_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x16 aom_highbd_12_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x32 aom_highbd_12_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x128 aom_highbd_12_variance128x128_sse2
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x64 aom_highbd_12_variance128x64_sse2
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_sse2
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_sse2
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x64 aom_highbd_12_variance16x64_sse2
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_sse2
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_sse2
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_sse2
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_sse2
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x8 aom_highbd_12_variance32x8_sse2
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x128 aom_highbd_12_variance64x128_sse2
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x16 aom_highbd_12_variance64x16_sse2
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_sse2
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_sse2
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_sse2
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x32 aom_highbd_12_variance8x32_sse2
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_sse2
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_sse2
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_sse2
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x128 aom_highbd_8_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x64 aom_highbd_8_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x16 aom_highbd_8_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x32 aom_highbd_8_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x4 aom_highbd_8_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x64 aom_highbd_8_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x8 aom_highbd_8_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x16 aom_highbd_8_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x32 aom_highbd_8_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x64 aom_highbd_8_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x8 aom_highbd_8_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x16 aom_highbd_8_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x4 aom_highbd_8_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x8 aom_highbd_8_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x128 aom_highbd_8_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x16 aom_highbd_8_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x32 aom_highbd_8_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x64 aom_highbd_8_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x16 aom_highbd_8_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x32 aom_highbd_8_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x4 aom_highbd_8_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x8 aom_highbd_8_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x128 aom_highbd_8_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x64 aom_highbd_8_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x4 aom_highbd_8_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x64 aom_highbd_8_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x8 aom_highbd_8_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x128 aom_highbd_8_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x16 aom_highbd_8_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x32 aom_highbd_8_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x128 aom_highbd_8_sub_pixel_variance128x128_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x64 aom_highbd_8_sub_pixel_variance128x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x4 aom_highbd_8_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x64 aom_highbd_8_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x8 aom_highbd_8_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x128 aom_highbd_8_sub_pixel_variance64x128_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x16 aom_highbd_8_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x32 aom_highbd_8_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x128 aom_highbd_8_variance128x128_sse2
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x64 aom_highbd_8_variance128x64_sse2
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_sse2
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_sse2
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x64 aom_highbd_8_variance16x64_sse2
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_sse2
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_sse2
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_sse2
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_sse2
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x8 aom_highbd_8_variance32x8_sse2
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x128 aom_highbd_8_variance64x128_sse2
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x16 aom_highbd_8_variance64x16_sse2
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_sse2
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_sse2
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_sse2
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x32 aom_highbd_8_variance8x32_sse2
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_sse2
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+#define aom_highbd_avg_4x4 aom_highbd_avg_4x4_c
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+#define aom_highbd_avg_8x8 aom_highbd_avg_8x8_c
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+#define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_highbd_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_sse2
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_sse2
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x4 aom_highbd_dc_128_predictor_16x4_c
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x64 aom_highbd_dc_128_predictor_16x64_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_sse2
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x64 aom_highbd_dc_128_predictor_32x64_c
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x8 aom_highbd_dc_128_predictor_32x8_c
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x16 aom_highbd_dc_128_predictor_4x16_c
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_sse2
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_sse2
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x16 aom_highbd_dc_128_predictor_64x16_c
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x32 aom_highbd_dc_128_predictor_64x32_c
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x64 aom_highbd_dc_128_predictor_64x64_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_sse2
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x32 aom_highbd_dc_128_predictor_8x32_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_sse2
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_sse2
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_sse2
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_sse2
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x4 aom_highbd_dc_left_predictor_16x4_c
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x64 aom_highbd_dc_left_predictor_16x64_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_sse2
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x64 aom_highbd_dc_left_predictor_32x64_c
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x8 aom_highbd_dc_left_predictor_32x8_c
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x16 aom_highbd_dc_left_predictor_4x16_c
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_sse2
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_sse2
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x16 aom_highbd_dc_left_predictor_64x16_c
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x32 aom_highbd_dc_left_predictor_64x32_c
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x64 aom_highbd_dc_left_predictor_64x64_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_sse2
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x32 aom_highbd_dc_left_predictor_8x32_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_sse2
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_sse2
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_sse2
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_sse2
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x4 aom_highbd_dc_predictor_16x4_c
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x64 aom_highbd_dc_predictor_16x64_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_sse2
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x64 aom_highbd_dc_predictor_32x64_c
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x8 aom_highbd_dc_predictor_32x8_c
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x16 aom_highbd_dc_predictor_4x16_c
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_sse2
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_sse2
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x16 aom_highbd_dc_predictor_64x16_c
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x32 aom_highbd_dc_predictor_64x32_c
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x64 aom_highbd_dc_predictor_64x64_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_sse2
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x32 aom_highbd_dc_predictor_8x32_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_sse2
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_sse2
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_sse2
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_sse2
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x4 aom_highbd_dc_top_predictor_16x4_c
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x64 aom_highbd_dc_top_predictor_16x64_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_sse2
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x64 aom_highbd_dc_top_predictor_32x64_c
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x8 aom_highbd_dc_top_predictor_32x8_c
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x16 aom_highbd_dc_top_predictor_4x16_c
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_sse2
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_sse2
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x16 aom_highbd_dc_top_predictor_64x16_c
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x32 aom_highbd_dc_top_predictor_64x32_c
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x64 aom_highbd_dc_top_predictor_64x64_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_sse2
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x32 aom_highbd_dc_top_predictor_8x32_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_sse2
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_sse2
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_highbd_dist_wtd_comp_avg_pred_sse2(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_highbd_dist_wtd_comp_avg_pred aom_highbd_dist_wtd_comp_avg_pred_sse2
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_sse2
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_sse2
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x4 aom_highbd_h_predictor_16x4_c
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x64 aom_highbd_h_predictor_16x64_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_sse2
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x64 aom_highbd_h_predictor_32x64_c
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x8 aom_highbd_h_predictor_32x8_c
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x16 aom_highbd_h_predictor_4x16_c
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_sse2
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_sse2
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x16 aom_highbd_h_predictor_64x16_c
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x32 aom_highbd_h_predictor_64x32_c
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x64 aom_highbd_h_predictor_64x64_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_sse2
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x32 aom_highbd_h_predictor_8x32_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_sse2
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_sse2
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_14 aom_highbd_lpf_horizontal_14_sse2
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_sse2
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_6 aom_highbd_lpf_horizontal_6_sse2
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_6_dual aom_highbd_lpf_horizontal_6_dual_sse2
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_8 aom_highbd_lpf_horizontal_8_sse2
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_14 aom_highbd_lpf_vertical_14_sse2
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_4 aom_highbd_lpf_vertical_4_sse2
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_6 aom_highbd_lpf_vertical_6_sse2
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_6_dual aom_highbd_lpf_vertical_6_dual_sse2
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_8 aom_highbd_lpf_vertical_8_sse2
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_64x64_adaptive aom_highbd_quantize_b_64x64_adaptive_sse2
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x16 aom_highbd_sad4x16_sse2
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x16_avg aom_highbd_sad4x16_avg_sse2
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x3d aom_highbd_sad4x16x3d_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x4d aom_highbd_sad4x16x4d_sse2
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x4 aom_highbd_sad4x4_sse2
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x4_avg aom_highbd_sad4x4_avg_sse2
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x3d aom_highbd_sad4x4x3d_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x4d aom_highbd_sad4x4x4d_sse2
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x8 aom_highbd_sad4x8_sse2
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x8_avg aom_highbd_sad4x8_avg_sse2
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x3d aom_highbd_sad4x8x3d_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x4d aom_highbd_sad4x8x4d_sse2
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x16 aom_highbd_sad8x16_sse2
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x16_avg aom_highbd_sad8x16_avg_sse2
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x3d aom_highbd_sad8x16x3d_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x4d aom_highbd_sad8x16x4d_sse2
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x32 aom_highbd_sad8x32_sse2
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x32_avg aom_highbd_sad8x32_avg_sse2
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x3d aom_highbd_sad8x32x3d_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x4d aom_highbd_sad8x32x4d_sse2
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x4 aom_highbd_sad8x4_sse2
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x4_avg aom_highbd_sad8x4_avg_sse2
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x3d aom_highbd_sad8x4x3d_c
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x4d aom_highbd_sad8x4x4d_sse2
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x8 aom_highbd_sad8x8_sse2
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x8_avg aom_highbd_sad8x8_avg_sse2
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x3d aom_highbd_sad8x8x3d_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x4d aom_highbd_sad8x8x4d_sse2
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x4 aom_highbd_sad_skip_16x4_c
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x4x4d aom_highbd_sad_skip_16x4x4d_c
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x16 aom_highbd_sad_skip_4x16_sse2
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x16x4d aom_highbd_sad_skip_4x16x4d_sse2
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x4 aom_highbd_sad_skip_4x4_c
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x4x4d aom_highbd_sad_skip_4x4x4d_c
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x8 aom_highbd_sad_skip_4x8_sse2
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x8x4d aom_highbd_sad_skip_4x8x4d_sse2
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x16 aom_highbd_sad_skip_8x16_sse2
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x16x4d aom_highbd_sad_skip_8x16x4d_sse2
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x32 aom_highbd_sad_skip_8x32_sse2
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x32x4d aom_highbd_sad_skip_8x32x4d_sse2
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x4 aom_highbd_sad_skip_8x4_c
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x4x4d aom_highbd_sad_skip_8x4x4d_c
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x8 aom_highbd_sad_skip_8x8_sse2
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x8x4d aom_highbd_sad_skip_8x8x4d_sse2
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_highbd_sse)(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define aom_highbd_subtract_block aom_highbd_subtract_block_sse2
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_sse2
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_sse2
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_sse2
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_sse2
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_sse2
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_sse2
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_sse2
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_sse2
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output);
+#define aom_ifft4x4_float aom_ifft4x4_float_sse2
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_avx2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_col)(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_avx2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_14 aom_lpf_horizontal_14_sse2
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_minmax_8x8 aom_minmax_8x8_sse2
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_sse2
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_sse2
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_sse2
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_16xh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_sse2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_avx2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit_highbd)(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32_adaptive aom_quantize_b_32x32_adaptive_sse2
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64_adaptive aom_quantize_b_64x64_adaptive_sse2
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x16 aom_sad16x16_sse2
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_sse2
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_sse2
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_sse2
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x4 aom_sad16x4_sse2
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x4_avg aom_sad16x4_avg_sse2
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x3d aom_sad16x4x3d_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x64 aom_sad16x64_sse2
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x64_avg aom_sad16x64_avg_sse2
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x8 aom_sad16x8_sse2
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_sse2
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x8 aom_sad32x8_sse2
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x8_avg aom_sad32x8_avg_sse2
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x16 aom_sad4x16_sse2
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x16_avg aom_sad4x16_avg_sse2
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x3d aom_sad4x16x3d_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x4d aom_sad4x16x4d_sse2
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x4 aom_sad4x4_sse2
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_sse2
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x3d aom_sad4x4x3d_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x4d aom_sad4x4x4d_sse2
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_sse2
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_sse2
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x3d aom_sad4x8x3d_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x4d aom_sad4x8x4d_sse2
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x16 aom_sad64x16_sse2
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x16_avg aom_sad64x16_avg_sse2
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x16 aom_sad8x16_sse2
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_sse2
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x3d aom_sad8x16x3d_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x4d aom_sad8x16x4d_sse2
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x32 aom_sad8x32_sse2
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x32_avg aom_sad8x32_avg_sse2
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x3d aom_sad8x32x3d_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x4d aom_sad8x32x4d_sse2
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_sse2
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_sse2
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x3d aom_sad8x4x3d_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x4d aom_sad8x4x4d_sse2
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x8 aom_sad8x8_sse2
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_sse2
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x3d aom_sad8x8x3d_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x4d aom_sad8x8x4d_sse2
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x16 aom_sad_skip_16x16_sse2
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x32 aom_sad_skip_16x32_sse2
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x4 aom_sad_skip_16x4_c
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x64 aom_sad_skip_16x64_sse2
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x8 aom_sad_skip_16x8_sse2
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x8 aom_sad_skip_32x8_sse2
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x16 aom_sad_skip_4x16_sse2
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x16x4d aom_sad_skip_4x16x4d_sse2
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x4 aom_sad_skip_4x4_c
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x4x4d aom_sad_skip_4x4x4d_c
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x8 aom_sad_skip_4x8_sse2
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x8x4d aom_sad_skip_4x8x4d_sse2
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x16 aom_sad_skip_64x16_sse2
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x16 aom_sad_skip_8x16_sse2
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x16x4d aom_sad_skip_8x16x4d_sse2
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x32 aom_sad_skip_8x32_sse2
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x32x4d aom_sad_skip_8x32x4d_sse2
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x4 aom_sad_skip_8x4_c
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x4x4d aom_sad_skip_8x4x4d_c
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x8 aom_sad_skip_8x8_sse2
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x8x4d aom_sad_skip_8x8x4d_sse2
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+int aom_satd_sse2(const tran_low_t *coeff, int length);
+int aom_satd_avx2(const tran_low_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const tran_low_t *coeff, int length);
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+int aom_satd_lp_sse2(const int16_t *coeff, int length);
+int aom_satd_lp_avx2(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd_lp)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_sse)(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+void aom_ssim_parms_8x8_sse2(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_sse2
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t N);
+#define aom_sum_squares_i16 aom_sum_squares_i16_sse2
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width, int height, int *sum);
+RTCD_EXTERN uint64_t (*aom_sum_sse_2d_i16)(const int16_t *src, int src_stride, int width, int height, int *sum);
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x16 aom_v_predictor_16x16_sse2
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x4 aom_v_predictor_4x4_sse2
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_sse2
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x8 aom_v_predictor_8x8_sse2
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u16)(uint8_t *src, int src_stride, int width, int height);
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u8)(uint8_t *src, int src_stride, int width, int height);
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x16 aom_variance4x16_sse2
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x4 aom_variance4x4_sse2
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x8 aom_variance4x8_sse2
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x16 aom_variance8x16_sse2
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x32 aom_variance8x32_sse2
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x4 aom_variance8x4_sse2
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x8 aom_variance8x8_sse2
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+
+void aom_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_avg_8x8_quad = aom_avg_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_avg_8x8_quad = aom_avg_8x8_quad_avx2;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
+    aom_comp_avg_pred = aom_comp_avg_pred_c;
+    if (flags & HAS_AVX2) aom_comp_avg_pred = aom_comp_avg_pred_avx2;
+    aom_comp_mask_pred = aom_comp_mask_pred_c;
+    if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
+    if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_flow_at_point = aom_compute_flow_at_point_c;
+    if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_horiz = aom_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_horiz = aom_convolve8_horiz_avx2;
+    aom_convolve8_vert = aom_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
+    aom_convolve_copy = aom_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_avx2;
+    aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
+    aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
+    aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_avx2;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_avx2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_avx2;
+    aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
+    aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
+    aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_avx2;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_avx2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_avx2;
+    aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
+    aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
+    aom_dc_predictor_64x64 = aom_dc_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_avx2;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_avx2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_avx2;
+    aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
+    aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
+    aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_avx2;
+    aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3;
+    aom_fft16x16_float = aom_fft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_fft16x16_float = aom_fft16x16_float_avx2;
+    aom_fft32x32_float = aom_fft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_fft32x32_float = aom_fft32x32_float_avx2;
+    aom_fft8x8_float = aom_fft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_fft8x8_float = aom_fft8x8_float_avx2;
+    aom_get_blk_sse_sum = aom_get_blk_sse_sum_sse2;
+    if (flags & HAS_AVX2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_avx2;
+    aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_avx2;
+    aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_avx2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_avx2;
+    aom_hadamard_16x16 = aom_hadamard_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_16x16 = aom_hadamard_16x16_avx2;
+    aom_hadamard_32x32 = aom_hadamard_32x32_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_32x32 = aom_hadamard_32x32_avx2;
+    aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_avx2;
+    aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_avx2;
+    aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
+    aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_sse4_1;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_sse4_1;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_sse4_1;
+    aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_sse4_1;
+    aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_sse4_1;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_sse4_1;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_sse4_1;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_sse4_1;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_sse4_1;
+    aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_sse4_1;
+    aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_sse4_1;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_sse4_1;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_sse4_1;
+    aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_sse4_1;
+    aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_sse4_1;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_sse4_1;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_sse4_1;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_sse4_1;
+    aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_sse4_1;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_sse4_1;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_avx2;
+    aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_avx2;
+    aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_avx2;
+    aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_avx2;
+    aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_avx2;
+    aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_avx2;
+    aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_avx2;
+    aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_avx2;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_avx2;
+    aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_avx2;
+    aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_avx2;
+    aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_avx2;
+    aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_avx2;
+    aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_avx2;
+    aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_avx2;
+    aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_avx2;
+    aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_avx2;
+    aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_avx2;
+    aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_avx2;
+    aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_avx2;
+    aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_avx2;
+    aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_avx2;
+    aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_avx2;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_sse4_1;
+    aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_avx2;
+    aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_avx2;
+    aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_avx2;
+    aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_avx2;
+    aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_avx2;
+    aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_avx2;
+    aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_avx2;
+    aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_sse4_1;
+    aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_sse4_1;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_sse4_1;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_sse4_1;
+    aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_sse4_1;
+    aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_sse4_1;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_sse4_1;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_sse4_1;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_sse4_1;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_sse4_1;
+    aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_sse4_1;
+    aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_sse4_1;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_sse4_1;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_sse4_1;
+    aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_sse4_1;
+    aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_sse4_1;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_sse4_1;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_sse4_1;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_sse4_1;
+    aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_sse4_1;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_sse4_1;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_sse4_1;
+    aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_sse4_1;
+    aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_sse4_1;
+    aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_sse4_1;
+    aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_sse4_1;
+    aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_sse4_1;
+    aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_sse4_1;
+    aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_sse4_1;
+    aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_sse4_1;
+    aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_sse4_1;
+    aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_sse4_1;
+    aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_sse4_1;
+    aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_sse4_1;
+    aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_sse4_1;
+    aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_sse4_1;
+    aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_sse4_1;
+    aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_sse4_1;
+    aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_sse4_1;
+    aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_sse4_1;
+    aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_sse4_1;
+    aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_sse4_1;
+    aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_sse4_1;
+    aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_sse4_1;
+    aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_avx2;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_sse4_1;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_sse4_1;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
+    aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_sse2;
+    if (flags & HAS_AVX2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_avx2;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_avx2;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_avx2;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_avx2;
+    aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_avx2;
+    aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_avx2;
+    aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_avx2;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_avx2;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_avx2;
+    aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_avx2;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_avx2;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
+    aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_avx2;
+    aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_avx2;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_avx2;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_avx2;
+    aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_avx2;
+    aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_avx2;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_avx2;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_avx2;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_avx2;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_avx2;
+    aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_avx2;
+    aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_avx2;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_avx2;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_avx2;
+    aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_avx2;
+    aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_avx2;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_avx2;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_avx2;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_avx2;
+    aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_avx2;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_avx2;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_avx2;
+    aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_avx2;
+    aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_avx2;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_avx2;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_avx2;
+    aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_avx2;
+    aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_avx2;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_avx2;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_avx2;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_avx2;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_avx2;
+    aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_avx2;
+    aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_avx2;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_avx2;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_avx2;
+    aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_avx2;
+    aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_avx2;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_avx2;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_avx2;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_avx2;
+    aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_avx2;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_avx2;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_avx2;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b = aom_highbd_quantize_b_avx2;
+    aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_avx2;
+    aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_avx2;
+    aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_avx2;
+    aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_avx2;
+    aom_highbd_sad128x128 = aom_highbd_sad128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128 = aom_highbd_sad128x128_avx2;
+    aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_avx2;
+    aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_avx2;
+    aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_avx2;
+    aom_highbd_sad128x64 = aom_highbd_sad128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64 = aom_highbd_sad128x64_avx2;
+    aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_avx2;
+    aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_avx2;
+    aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_avx2;
+    aom_highbd_sad16x16 = aom_highbd_sad16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16 = aom_highbd_sad16x16_avx2;
+    aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_avx2;
+    aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_avx2;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_avx2;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32 = aom_highbd_sad16x32_avx2;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_avx2;
+    aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_avx2;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_avx2;
+    aom_highbd_sad16x4 = aom_highbd_sad16x4_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4 = aom_highbd_sad16x4_avx2;
+    aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_avx2;
+    aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_avx2;
+    aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_avx2;
+    aom_highbd_sad16x64 = aom_highbd_sad16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64 = aom_highbd_sad16x64_avx2;
+    aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_avx2;
+    aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_avx2;
+    aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_avx2;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8 = aom_highbd_sad16x8_avx2;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_avx2;
+    aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_avx2;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_avx2;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16 = aom_highbd_sad32x16_avx2;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_avx2;
+    aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_avx2;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_avx2;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32 = aom_highbd_sad32x32_avx2;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_avx2;
+    aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_avx2;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_avx2;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64 = aom_highbd_sad32x64_avx2;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_avx2;
+    aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_avx2;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_avx2;
+    aom_highbd_sad32x8 = aom_highbd_sad32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8 = aom_highbd_sad32x8_avx2;
+    aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_avx2;
+    aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_avx2;
+    aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_avx2;
+    aom_highbd_sad64x128 = aom_highbd_sad64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128 = aom_highbd_sad64x128_avx2;
+    aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_avx2;
+    aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_avx2;
+    aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_avx2;
+    aom_highbd_sad64x16 = aom_highbd_sad64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16 = aom_highbd_sad64x16_avx2;
+    aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_avx2;
+    aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_avx2;
+    aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_avx2;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32 = aom_highbd_sad64x32_avx2;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_avx2;
+    aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_avx2;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_avx2;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64 = aom_highbd_sad64x64_avx2;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_avx2;
+    aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_avx2;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_avx2;
+    aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_avx2;
+    aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_avx2;
+    aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_avx2;
+    aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_avx2;
+    aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_avx2;
+    aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_avx2;
+    aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_avx2;
+    aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_avx2;
+    aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_avx2;
+    aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_avx2;
+    aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_avx2;
+    aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_avx2;
+    aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_avx2;
+    aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_avx2;
+    aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_avx2;
+    aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_avx2;
+    aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_avx2;
+    aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_avx2;
+    aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_avx2;
+    aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_avx2;
+    aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_avx2;
+    aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_avx2;
+    aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_avx2;
+    aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_avx2;
+    aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_avx2;
+    aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_avx2;
+    aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_avx2;
+    aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_avx2;
+    aom_highbd_sse = aom_highbd_sse_c;
+    if (flags & HAS_SSE4_1) aom_highbd_sse = aom_highbd_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_sse = aom_highbd_sse_avx2;
+    aom_ifft16x16_float = aom_ifft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft16x16_float = aom_ifft16x16_float_avx2;
+    aom_ifft32x32_float = aom_ifft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft32x32_float = aom_ifft32x32_float_avx2;
+    aom_ifft8x8_float = aom_ifft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft8x8_float = aom_ifft8x8_float_avx2;
+    aom_int_pro_col = aom_int_pro_col_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_col = aom_int_pro_col_avx2;
+    aom_int_pro_row = aom_int_pro_row_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_row = aom_int_pro_row_avx2;
+    aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2;
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2;
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2;
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_avx2;
+    aom_masked_sad128x128 = aom_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128 = aom_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x128 = aom_masked_sad128x128_avx2;
+    aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_ssse3;
+    aom_masked_sad128x64 = aom_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64 = aom_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x64 = aom_masked_sad128x64_avx2;
+    aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_ssse3;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x16 = aom_masked_sad16x16_avx2;
+    aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_ssse3;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32 = aom_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x32 = aom_masked_sad16x32_avx2;
+    aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_ssse3;
+    aom_masked_sad16x4 = aom_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4 = aom_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x4 = aom_masked_sad16x4_avx2;
+    aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_ssse3;
+    aom_masked_sad16x64 = aom_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64 = aom_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x64 = aom_masked_sad16x64_avx2;
+    aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_ssse3;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8 = aom_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x8 = aom_masked_sad16x8_avx2;
+    aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_ssse3;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16 = aom_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x16 = aom_masked_sad32x16_avx2;
+    aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_ssse3;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32 = aom_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x32 = aom_masked_sad32x32_avx2;
+    aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_ssse3;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64 = aom_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x64 = aom_masked_sad32x64_avx2;
+    aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_ssse3;
+    aom_masked_sad32x8 = aom_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8 = aom_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x8 = aom_masked_sad32x8_avx2;
+    aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_ssse3;
+    aom_masked_sad4x16 = aom_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16 = aom_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x16 = aom_masked_sad4x16_avx2;
+    aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_ssse3;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4 = aom_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x4 = aom_masked_sad4x4_avx2;
+    aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_ssse3;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8 = aom_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x8 = aom_masked_sad4x8_avx2;
+    aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_ssse3;
+    aom_masked_sad64x128 = aom_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128 = aom_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x128 = aom_masked_sad64x128_avx2;
+    aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_ssse3;
+    aom_masked_sad64x16 = aom_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16 = aom_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x16 = aom_masked_sad64x16_avx2;
+    aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_ssse3;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32 = aom_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x32 = aom_masked_sad64x32_avx2;
+    aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_ssse3;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64 = aom_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x64 = aom_masked_sad64x64_avx2;
+    aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_ssse3;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16 = aom_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x16 = aom_masked_sad8x16_avx2;
+    aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_ssse3;
+    aom_masked_sad8x32 = aom_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32 = aom_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x32 = aom_masked_sad8x32_avx2;
+    aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_ssse3;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4 = aom_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x4 = aom_masked_sad8x4_avx2;
+    aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_ssse3;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8 = aom_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x8 = aom_masked_sad8x8_avx2;
+    aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_ssse3;
+    aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_ssse3;
+    aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_ssse3;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_ssse3;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_ssse3;
+    aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_ssse3;
+    aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_ssse3;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_ssse3;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_ssse3;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_ssse3;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_ssse3;
+    aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_ssse3;
+    aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_ssse3;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_ssse3;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_ssse3;
+    aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_ssse3;
+    aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_ssse3;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_ssse3;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_ssse3;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_ssse3;
+    aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_ssse3;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_ssse3;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_ssse3;
+    aom_mse16x16 = aom_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
+    aom_mse_16xh_16bit = aom_mse_16xh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_avx2;
+    aom_mse_wxh_16bit = aom_mse_wxh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_avx2;
+    aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_avx2;
+    aom_obmc_sad128x128 = aom_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x128 = aom_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x128 = aom_obmc_sad128x128_avx2;
+    aom_obmc_sad128x64 = aom_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x64 = aom_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x64 = aom_obmc_sad128x64_avx2;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x16 = aom_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x16 = aom_obmc_sad16x16_avx2;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x32 = aom_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x32 = aom_obmc_sad16x32_avx2;
+    aom_obmc_sad16x4 = aom_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x4 = aom_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x4 = aom_obmc_sad16x4_avx2;
+    aom_obmc_sad16x64 = aom_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x64 = aom_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x64 = aom_obmc_sad16x64_avx2;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x8 = aom_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x8 = aom_obmc_sad16x8_avx2;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x16 = aom_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x16 = aom_obmc_sad32x16_avx2;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x32 = aom_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x32 = aom_obmc_sad32x32_avx2;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x64 = aom_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x64 = aom_obmc_sad32x64_avx2;
+    aom_obmc_sad32x8 = aom_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x8 = aom_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x8 = aom_obmc_sad32x8_avx2;
+    aom_obmc_sad4x16 = aom_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x16 = aom_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x16 = aom_obmc_sad4x16_avx2;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x4 = aom_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x4 = aom_obmc_sad4x4_avx2;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x8 = aom_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x8 = aom_obmc_sad4x8_avx2;
+    aom_obmc_sad64x128 = aom_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x128 = aom_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x128 = aom_obmc_sad64x128_avx2;
+    aom_obmc_sad64x16 = aom_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x16 = aom_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x16 = aom_obmc_sad64x16_avx2;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x32 = aom_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x32 = aom_obmc_sad64x32_avx2;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x64 = aom_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x64 = aom_obmc_sad64x64_avx2;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x16 = aom_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x16 = aom_obmc_sad8x16_avx2;
+    aom_obmc_sad8x32 = aom_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x32 = aom_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x32 = aom_obmc_sad8x32_avx2;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x4 = aom_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x4 = aom_obmc_sad8x4_avx2;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x8 = aom_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x8 = aom_obmc_sad8x8_avx2;
+    aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_sse4_1;
+    aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_sse4_1;
+    aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_sse4_1;
+    aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_sse4_1;
+    aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_sse4_1;
+    aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_sse4_1;
+    aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_sse4_1;
+    aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_sse4_1;
+    aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_sse4_1;
+    aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_sse4_1;
+    aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_sse4_1;
+    aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_sse4_1;
+    aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_sse4_1;
+    aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_sse4_1;
+    aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_sse4_1;
+    aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_sse4_1;
+    aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_sse4_1;
+    aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_sse4_1;
+    aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_sse4_1;
+    aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_sse4_1;
+    aom_obmc_variance128x128 = aom_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x128 = aom_obmc_variance128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x128 = aom_obmc_variance128x128_avx2;
+    aom_obmc_variance128x64 = aom_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x64 = aom_obmc_variance128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x64 = aom_obmc_variance128x64_avx2;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x16 = aom_obmc_variance16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x16 = aom_obmc_variance16x16_avx2;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x32 = aom_obmc_variance16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x32 = aom_obmc_variance16x32_avx2;
+    aom_obmc_variance16x4 = aom_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x4 = aom_obmc_variance16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x4 = aom_obmc_variance16x4_avx2;
+    aom_obmc_variance16x64 = aom_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x64 = aom_obmc_variance16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x64 = aom_obmc_variance16x64_avx2;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x8 = aom_obmc_variance16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x8 = aom_obmc_variance16x8_avx2;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x16 = aom_obmc_variance32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x16 = aom_obmc_variance32x16_avx2;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x32 = aom_obmc_variance32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x32 = aom_obmc_variance32x32_avx2;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x64 = aom_obmc_variance32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x64 = aom_obmc_variance32x64_avx2;
+    aom_obmc_variance32x8 = aom_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x8 = aom_obmc_variance32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x8 = aom_obmc_variance32x8_avx2;
+    aom_obmc_variance4x16 = aom_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x16 = aom_obmc_variance4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x16 = aom_obmc_variance4x16_avx2;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x4 = aom_obmc_variance4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x4 = aom_obmc_variance4x4_avx2;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x8 = aom_obmc_variance4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x8 = aom_obmc_variance4x8_avx2;
+    aom_obmc_variance64x128 = aom_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x128 = aom_obmc_variance64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x128 = aom_obmc_variance64x128_avx2;
+    aom_obmc_variance64x16 = aom_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x16 = aom_obmc_variance64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x16 = aom_obmc_variance64x16_avx2;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x32 = aom_obmc_variance64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x32 = aom_obmc_variance64x32_avx2;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x64 = aom_obmc_variance64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x64 = aom_obmc_variance64x64_avx2;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x16 = aom_obmc_variance8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x16 = aom_obmc_variance8x16_avx2;
+    aom_obmc_variance8x32 = aom_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x32 = aom_obmc_variance8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x32 = aom_obmc_variance8x32_avx2;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x4 = aom_obmc_variance8x4_avx2;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x8 = aom_obmc_variance8x8_avx2;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_avx2;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_avx2;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_avx2;
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_avx2;
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_avx2;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3;
+    aom_quantize_b = aom_quantize_b_sse2;
+    if (flags & HAS_SSSE3) aom_quantize_b = aom_quantize_b_ssse3;
+    if (flags & HAS_AVX) aom_quantize_b = aom_quantize_b_avx;
+    if (flags & HAS_AVX2) aom_quantize_b = aom_quantize_b_avx2;
+    aom_quantize_b_32x32 = aom_quantize_b_32x32_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_32x32 = aom_quantize_b_32x32_ssse3;
+    if (flags & HAS_AVX) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx;
+    if (flags & HAS_AVX2) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx2;
+    aom_quantize_b_64x64 = aom_quantize_b_64x64_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_64x64 = aom_quantize_b_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_quantize_b_64x64 = aom_quantize_b_64x64_avx2;
+    aom_quantize_b_adaptive = aom_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_avx2;
+    aom_sad128x128 = aom_sad128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128 = aom_sad128x128_avx2;
+    aom_sad128x128_avg = aom_sad128x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128_avg = aom_sad128x128_avg_avx2;
+    aom_sad128x128x3d = aom_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x128x3d = aom_sad128x128x3d_avx2;
+    aom_sad128x128x4d = aom_sad128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128x4d = aom_sad128x128x4d_avx2;
+    aom_sad128x64 = aom_sad128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64 = aom_sad128x64_avx2;
+    aom_sad128x64_avg = aom_sad128x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64_avg = aom_sad128x64_avg_avx2;
+    aom_sad128x64x3d = aom_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x64x3d = aom_sad128x64x3d_avx2;
+    aom_sad128x64x4d = aom_sad128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64x4d = aom_sad128x64x4d_avx2;
+    aom_sad16x16x3d = aom_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x16x3d = aom_sad16x16x3d_avx2;
+    aom_sad16x16x4d = aom_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x16x4d = aom_sad16x16x4d_avx2;
+    aom_sad16x32x3d = aom_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
+    aom_sad16x32x4d = aom_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4x4d = aom_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
+    aom_sad16x64x3d = aom_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x64x3d = aom_sad16x64x3d_avx2;
+    aom_sad16x64x4d = aom_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x64x4d = aom_sad16x64x4d_avx2;
+    aom_sad16x8x3d = aom_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x8x3d = aom_sad16x8x3d_avx2;
+    aom_sad16x8x4d = aom_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x8x4d = aom_sad16x8x4d_avx2;
+    aom_sad32x16 = aom_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16 = aom_sad32x16_avx2;
+    aom_sad32x16_avg = aom_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16_avg = aom_sad32x16_avg_avx2;
+    aom_sad32x16x3d = aom_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x16x3d = aom_sad32x16x3d_avx2;
+    aom_sad32x16x4d = aom_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16x4d = aom_sad32x16x4d_avx2;
+    aom_sad32x32 = aom_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32 = aom_sad32x32_avx2;
+    aom_sad32x32_avg = aom_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32_avg = aom_sad32x32_avg_avx2;
+    aom_sad32x32x3d = aom_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x32x3d = aom_sad32x32x3d_avx2;
+    aom_sad32x32x4d = aom_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32x4d = aom_sad32x32x4d_avx2;
+    aom_sad32x64 = aom_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64 = aom_sad32x64_avx2;
+    aom_sad32x64_avg = aom_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64_avg = aom_sad32x64_avg_avx2;
+    aom_sad32x64x3d = aom_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x64x3d = aom_sad32x64x3d_avx2;
+    aom_sad32x64x4d = aom_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64x4d = aom_sad32x64x4d_avx2;
+    aom_sad32x8x3d = aom_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x8x3d = aom_sad32x8x3d_avx2;
+    aom_sad32x8x4d = aom_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x8x4d = aom_sad32x8x4d_avx2;
+    aom_sad64x128 = aom_sad64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128 = aom_sad64x128_avx2;
+    aom_sad64x128_avg = aom_sad64x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128_avg = aom_sad64x128_avg_avx2;
+    aom_sad64x128x3d = aom_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x128x3d = aom_sad64x128x3d_avx2;
+    aom_sad64x128x4d = aom_sad64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128x4d = aom_sad64x128x4d_avx2;
+    aom_sad64x16x3d = aom_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x16x3d = aom_sad64x16x3d_avx2;
+    aom_sad64x16x4d = aom_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x16x4d = aom_sad64x16x4d_avx2;
+    aom_sad64x32 = aom_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32 = aom_sad64x32_avx2;
+    aom_sad64x32_avg = aom_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32_avg = aom_sad64x32_avg_avx2;
+    aom_sad64x32x3d = aom_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x32x3d = aom_sad64x32x3d_avx2;
+    aom_sad64x32x4d = aom_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32x4d = aom_sad64x32x4d_avx2;
+    aom_sad64x64 = aom_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64 = aom_sad64x64_avx2;
+    aom_sad64x64_avg = aom_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64_avg = aom_sad64x64_avg_avx2;
+    aom_sad64x64x3d = aom_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x64x3d = aom_sad64x64x3d_avx2;
+    aom_sad64x64x4d = aom_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64x4d = aom_sad64x64x4d_avx2;
+    aom_sad_skip_128x128 = aom_sad_skip_128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128 = aom_sad_skip_128x128_avx2;
+    aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_avx2;
+    aom_sad_skip_128x64 = aom_sad_skip_128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64 = aom_sad_skip_128x64_avx2;
+    aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_avx2;
+    aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
+    aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
+    aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_avx2;
+    aom_sad_skip_32x16 = aom_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16 = aom_sad_skip_32x16_avx2;
+    aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_avx2;
+    aom_sad_skip_32x32 = aom_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32 = aom_sad_skip_32x32_avx2;
+    aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_avx2;
+    aom_sad_skip_32x64 = aom_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64 = aom_sad_skip_32x64_avx2;
+    aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_avx2;
+    aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_avx2;
+    aom_sad_skip_64x128 = aom_sad_skip_64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128 = aom_sad_skip_64x128_avx2;
+    aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_avx2;
+    aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_avx2;
+    aom_sad_skip_64x32 = aom_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32 = aom_sad_skip_64x32_avx2;
+    aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_avx2;
+    aom_sad_skip_64x64 = aom_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64 = aom_sad_skip_64x64_avx2;
+    aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_avx2;
+    aom_satd = aom_satd_sse2;
+    if (flags & HAS_AVX2) aom_satd = aom_satd_avx2;
+    aom_satd_lp = aom_satd_lp_sse2;
+    if (flags & HAS_AVX2) aom_satd_lp = aom_satd_lp_avx2;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3;
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_ssse3;
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_ssse3;
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_ssse3;
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_ssse3;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_ssse3;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_ssse3;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_ssse3;
+    aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
+    aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
+    aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_ssse3;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_ssse3;
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_ssse3;
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_ssse3;
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_ssse3;
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_ssse3;
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_ssse3;
+    aom_sse = aom_sse_c;
+    if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_subtract_block = aom_subtract_block_sse2;
+    if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_avx2;
+    aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_avx2;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x16 = aom_v_predictor_32x16_avx2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x32 = aom_v_predictor_32x32_avx2;
+    aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
+    aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
+    aom_v_predictor_64x64 = aom_v_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x64 = aom_v_predictor_64x64_avx2;
+    aom_var_2d_u16 = aom_var_2d_u16_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u16 = aom_var_2d_u16_avx2;
+    aom_var_2d_u8 = aom_var_2d_u8_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u8 = aom_var_2d_u8_avx2;
+    aom_variance128x128 = aom_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_variance128x128 = aom_variance128x128_avx2;
+    aom_variance128x64 = aom_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_variance128x64 = aom_variance128x64_avx2;
+    aom_variance16x16 = aom_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
+    aom_variance16x32 = aom_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_variance16x32 = aom_variance16x32_avx2;
+    aom_variance16x4 = aom_variance16x4_sse2;
+    if (flags & HAS_AVX2) aom_variance16x4 = aom_variance16x4_avx2;
+    aom_variance16x64 = aom_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_variance16x64 = aom_variance16x64_avx2;
+    aom_variance16x8 = aom_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_variance16x8 = aom_variance16x8_avx2;
+    aom_variance32x16 = aom_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_variance32x16 = aom_variance32x16_avx2;
+    aom_variance32x32 = aom_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_variance32x32 = aom_variance32x32_avx2;
+    aom_variance32x64 = aom_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_variance32x64 = aom_variance32x64_avx2;
+    aom_variance32x8 = aom_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_variance32x8 = aom_variance32x8_avx2;
+    aom_variance64x128 = aom_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_variance64x128 = aom_variance64x128_avx2;
+    aom_variance64x16 = aom_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_variance64x16 = aom_variance64x16_avx2;
+    aom_variance64x32 = aom_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_variance64x32 = aom_variance64x32_avx2;
+    aom_variance64x64 = aom_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_variance64x64 = aom_variance64x64_avx2;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
+    if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
+    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
+    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/x64/config/aom_scale_rtcd.h b/media/libaom/config/linux/x64/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..3b70fb47c3
--- /dev/null
+++ b/media/libaom/config/linux/x64/config/aom_scale_rtcd.h
@@ -0,0 +1,103 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/linux/x64/config/av1_rtcd.h b/media/libaom/config/linux/x64/config/av1_rtcd.h
new file mode 100644
index 0000000000..b1cdc99700
--- /dev/null
+++ b/media/libaom/config/linux/x64/config/av1_rtcd.h
@@ -0,0 +1,1217 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+void aom_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_sse2
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_dist_wtd_comp_avg_upsampled_pred_ssse3(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_sse2
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+#define aom_highbd_dist_wtd_comp_avg_upsampled_pred aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_upsampled_pred aom_highbd_upsampled_pred_sse2
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+#define aom_quantize_b_helper aom_quantize_b_helper_c
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+#define aom_upsampled_pred aom_upsampled_pred_sse2
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_sse4_1(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_avx2(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_avx2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+RTCD_EXTERN int64_t (*av1_block_error_lp)(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim1)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim2)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params_high_bd)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+void av1_cnn_convolve_no_maxpool_padding_valid_avx2(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+RTCD_EXTERN void (*av1_cnn_convolve_no_maxpool_padding_valid)(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+RTCD_EXTERN void (*av1_compute_stats)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+RTCD_EXTERN void (*av1_compute_stats_highbd)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr_intrabc av1_convolve_2d_sr_intrabc_c
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr_intrabc av1_convolve_x_sr_intrabc_c
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+double av1_estimate_noise_from_single_plane_avx2(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+RTCD_EXTERN double (*av1_estimate_noise_from_single_plane)(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void av1_fwht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*av1_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+uint32_t av1_get_crc32c_value_sse4_2(void *crc_calculator, uint8_t *p, size_t length);
+RTCD_EXTERN uint32_t (*av1_get_crc32c_value)(void *crc_calculator, uint8_t *p, size_t length);
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_sse4_1(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_avx2(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+RTCD_EXTERN void (*av1_get_horver_correlation_full)(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+void av1_get_nz_map_contexts_sse2(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+#define av1_get_nz_map_contexts av1_get_nz_map_contexts_sse2
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_highbd_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+RTCD_EXTERN int64_t (*av1_highbd_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_sse2
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_sse2
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr_intrabc av1_highbd_convolve_2d_sr_intrabc_c
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr_intrabc av1_highbd_convolve_x_sr_intrabc_c
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr_intrabc av1_highbd_convolve_y_sr_intrabc_c
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+#define av1_highbd_estimate_noise_from_single_plane av1_highbd_estimate_noise_from_single_plane_c
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+void av1_highbd_filter_intra_edge_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_highbd_filter_intra_edge)(uint16_t *p, int sz, int strength);
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_highbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_sse4_1(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+RTCD_EXTERN void (*av1_highbd_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+void av1_highbd_upsample_intra_edge_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_highbd_upsample_intra_edge)(uint16_t *p, int sz, int bd);
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_lowbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+void av1_nn_fast_softmax_16_sse3(const float *input_nodes, float *output);
+RTCD_EXTERN void (*av1_nn_fast_softmax_16)(const float *input_nodes, float *output);
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_sse3(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_avx2(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+RTCD_EXTERN void (*av1_nn_predict)(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_avx2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_lp)(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit);
+RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit);
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+RTCD_EXTERN void (*av1_txb_init_levels)(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_avx2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+RTCD_EXTERN void (*av1_wedge_compute_delta_squares)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_avx2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+RTCD_EXTERN int8_t (*av1_wedge_sign_from_residuals)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+RTCD_EXTERN uint64_t (*av1_wedge_sse_from_residuals)(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*cfl_get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*cfl_get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*cfl_get_subtract_average_fn)(TX_SIZE tx_size);
+
+void av1_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
+    av1_apply_temporal_filter = av1_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_apply_temporal_filter = av1_apply_temporal_filter_avx2;
+    av1_block_error = av1_block_error_sse2;
+    if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_block_error_lp = av1_block_error_lp_sse2;
+    if (flags & HAS_AVX2) av1_block_error_lp = av1_block_error_lp_avx2;
+    av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
+    av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
+    av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
+    if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
+    av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
+    av1_calc_indices_dim2 = av1_calc_indices_dim2_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim2 = av1_calc_indices_dim2_avx2;
+    av1_calc_proj_params = av1_calc_proj_params_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params = av1_calc_proj_params_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params = av1_calc_proj_params_avx2;
+    av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_avx2;
+    av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_c;
+    if (flags & HAS_AVX2) av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_avx2;
+    av1_compute_stats = av1_compute_stats_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats = av1_compute_stats_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats = av1_compute_stats_avx2;
+    av1_compute_stats_highbd = av1_compute_stats_highbd_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats_highbd = av1_compute_stats_highbd_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats_highbd = av1_compute_stats_highbd_avx2;
+    av1_convolve_2d_scale = av1_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
+    av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
+    av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
+    av1_convolve_x_sr = av1_convolve_x_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
+    av1_convolve_y_sr = av1_convolve_y_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
+    av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_avx2;
+    av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_avx2;
+    av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2;
+    av1_dr_prediction_z1 = av1_dr_prediction_z1_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2;
+    av1_dr_prediction_z2 = av1_dr_prediction_z2_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2;
+    av1_dr_prediction_z3 = av1_dr_prediction_z3_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2;
+    av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_c;
+    if (flags & HAS_AVX2) av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_avx2;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
+    av1_filter_intra_predictor = av1_filter_intra_predictor_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_avx2;
+    av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_sse4_1;
+    av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_sse4_1;
+    av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_sse4_1;
+    av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_avx2;
+    av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_sse4_1;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_avx2;
+    av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_sse4_1;
+    av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_sse4_1;
+    av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_sse4_1;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
+    av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_sse4_1;
+    av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_sse4_1;
+    av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_sse4_1;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_avx2;
+    av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_avx2;
+    av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_sse4_1;
+    av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_sse4_1;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_avx2;
+    av1_fwht4x4 = av1_fwht4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwht4x4 = av1_fwht4x4_sse4_1;
+    av1_get_crc32c_value = av1_get_crc32c_value_c;
+    if (flags & HAS_SSE4_2) av1_get_crc32c_value = av1_get_crc32c_value_sse4_2;
+    av1_get_horver_correlation_full = av1_get_horver_correlation_full_c;
+    if (flags & HAS_SSE4_1) av1_get_horver_correlation_full = av1_get_horver_correlation_full_sse4_1;
+    if (flags & HAS_AVX2) av1_get_horver_correlation_full = av1_get_horver_correlation_full_avx2;
+    av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_avx2;
+    av1_highbd_block_error = av1_highbd_block_error_sse2;
+    if (flags & HAS_AVX2) av1_highbd_block_error = av1_highbd_block_error_avx2;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
+    av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
+    av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
+    av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
+    av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_avx2;
+    av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_avx2;
+    av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_avx2;
+    av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_avx2;
+    av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_avx2;
+    av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_avx2;
+    av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_avx2;
+    av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_sse4_1;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_sse4_1;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_sse4_1;
+    av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
+    av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1;
+    av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_avx2;
+    av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
+    if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
+    av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_sse4_1;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2;
+    av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
+    if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
+    av1_inv_txfm_add = av1_inv_txfm_add_c;
+    if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
+    if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
+    av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_avx2;
+    av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_c;
+    if (flags & HAS_SSE3) av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_sse3;
+    av1_nn_predict = av1_nn_predict_c;
+    if (flags & HAS_SSE3) av1_nn_predict = av1_nn_predict_sse3;
+    if (flags & HAS_AVX2) av1_nn_predict = av1_nn_predict_avx2;
+    av1_quantize_fp = av1_quantize_fp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
+    av1_quantize_fp_64x64 = av1_quantize_fp_64x64_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_64x64 = av1_quantize_fp_64x64_avx2;
+    av1_quantize_lp = av1_quantize_lp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_lp = av1_quantize_lp_avx2;
+    av1_resize_and_extend_frame = av1_resize_and_extend_frame_c;
+    if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
+    av1_txb_init_levels = av1_txb_init_levels_c;
+    if (flags & HAS_SSE4_1) av1_txb_init_levels = av1_txb_init_levels_sse4_1;
+    if (flags & HAS_AVX2) av1_txb_init_levels = av1_txb_init_levels_avx2;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_warp_affine = av1_warp_affine_avx2;
+    av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
+    if (flags & HAS_AVX2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
+    av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_avx2;
+    av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
+    av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
+    if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
+    cdef_filter_16_0 = cdef_filter_16_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
+    cdef_filter_16_1 = cdef_filter_16_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
+    cdef_filter_16_2 = cdef_filter_16_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
+    cdef_filter_16_3 = cdef_filter_16_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
+    cdef_filter_8_0 = cdef_filter_8_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
+    cdef_filter_8_1 = cdef_filter_8_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
+    cdef_filter_8_2 = cdef_filter_8_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
+    cdef_filter_8_3 = cdef_filter_8_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
+    cdef_find_dir = cdef_find_dir_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
+    cdef_find_dir_dual = cdef_find_dir_dual_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
+    cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
+    cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
+    cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
+    cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
+    cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
+    cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
+    cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_avx2;
+    cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_avx2;
+    cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/mac/x64/config/aom_config.asm b/media/libaom/config/mac/x64/config/aom_config.asm
new file mode 100644
index 0000000000..f793ff3c6d
--- /dev/null
+++ b/media/libaom/config/mac/x64/config/aom_config.asm
@@ -0,0 +1,92 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+AOM_ARCH_AARCH64 equ 0
+AOM_ARCH_ARM equ 0
+AOM_ARCH_PPC equ 0
+AOM_ARCH_X86 equ 0
+AOM_ARCH_X86_64 equ 1
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_HIGHBITDEPTH equ 1
+CONFIG_AV1_TEMPORAL_DENOISING equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITRATE_ACCURACY equ 0
+CONFIG_BITRATE_ACCURACY_BL equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_COMPONENT_TIMING equ 0
+CONFIG_COLLECT_PARTITION_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_CWG_C013 equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_EXCLUDE_SIMD_MISMATCH equ 0
+CONFIG_FPMT_TEST equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBVMAF_PSNR_PEAK equ 1
+CONFIG_LIBYUV equ 0
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NN_V2 equ 0
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OPTICAL_FLOW_API equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_OUTPUT_FRAME_SIZE equ 0
+CONFIG_PARTITION_SEARCH_ORDER equ 0
+CONFIG_PIC equ 0
+CONFIG_RATECTRL_LOG equ 0
+CONFIG_RD_COMMAND equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_RT_ML_PARTITIONING equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SALIENCY_MAP equ 0
+CONFIG_SHARED equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_SPEED_STATS equ 0
+CONFIG_TFLITE equ 0
+CONFIG_THREE_PASS equ 0
+CONFIG_TUNE_BUTTERAUGLI equ 0
+CONFIG_TUNE_VMAF equ 0
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+FORCE_HIGHBITDEPTH_DECODING equ 0
+HAVE_ARM_CRC32 equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_FEXCEPT equ 1
+HAVE_MMX equ 1
+HAVE_NEON equ 0
+HAVE_NEON_DOTPROD equ 0
+HAVE_NEON_I8MM equ 0
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 1
+HAVE_SSE4_1 equ 1
+HAVE_SSE4_2 equ 1
+HAVE_SSSE3 equ 1
+HAVE_SVE equ 0
+HAVE_VSX equ 0
+HAVE_WXWIDGETS equ 0
+STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/mac/x64/config/aom_config.h b/media/libaom/config/mac/x64/config/aom_config.h
new file mode 100644
index 0000000000..670d2ffe56
--- /dev/null
+++ b/media/libaom/config/mac/x64/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 0
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 0
+#define AOM_ARCH_X86_64 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 0
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 1
+#define HAVE_NEON 0
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_SSE4_2 1
+#define HAVE_SSSE3 1
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h b/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..8e979cc189
--- /dev/null
+++ b/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h
@@ -0,0 +1,8036 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+#define aom_avg_4x4 aom_avg_4x4_sse2
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+#define aom_avg_8x8 aom_avg_8x8_sse2
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_avx2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_sse2
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_sse2
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_sse2
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_sse2
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_sse2
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_sse2
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_sse2
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_sse2
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x16 aom_dc_predictor_16x16_sse2
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x4 aom_dc_predictor_4x4_sse2
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_sse2
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x8 aom_dc_predictor_8x8_sse2
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_sse2
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_sse2
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_sse2
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_sse2
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_dist_wtd_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x128_avg aom_dist_wtd_sad128x128_avg_sse2
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x64_avg aom_dist_wtd_sad128x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x16_avg aom_dist_wtd_sad16x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x32_avg aom_dist_wtd_sad16x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x4_avg aom_dist_wtd_sad16x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x64_avg aom_dist_wtd_sad16x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x8_avg aom_dist_wtd_sad16x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x16_avg aom_dist_wtd_sad32x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x32_avg aom_dist_wtd_sad32x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x64_avg aom_dist_wtd_sad32x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x8_avg aom_dist_wtd_sad32x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x16_avg aom_dist_wtd_sad4x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x4_avg aom_dist_wtd_sad4x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x8_avg aom_dist_wtd_sad4x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x128_avg aom_dist_wtd_sad64x128_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x16_avg aom_dist_wtd_sad64x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x32_avg aom_dist_wtd_sad64x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x64_avg aom_dist_wtd_sad64x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x16_avg aom_dist_wtd_sad8x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x32_avg aom_dist_wtd_sad8x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x4_avg aom_dist_wtd_sad8x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x8_avg aom_dist_wtd_sad8x8_avg_sse2
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_sse2
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+void aom_fdct4x4_lp_sse2(const int16_t *input, int16_t *output, int stride);
+#define aom_fdct4x4_lp aom_fdct4x4_lp_sse2
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+void aom_fft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_fft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+void aom_fft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_fft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+void aom_fft4x4_float_sse2(const float *input, float *temp, float *output);
+#define aom_fft4x4_float aom_fft4x4_float_sse2
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+void aom_fft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_fft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_sse2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_avx2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+RTCD_EXTERN void (*aom_get_blk_sse_sum)(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_sse2
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+RTCD_EXTERN void (*aom_get_var_sse_sum_16x16_dual)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+RTCD_EXTERN void (*aom_get_var_sse_sum_8x8_quad)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x16 aom_h_predictor_16x16_sse2
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x4 aom_h_predictor_4x4_sse2
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x32 aom_h_predictor_64x32_sse2
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x64 aom_h_predictor_64x64_sse2
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_sse2
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x8 aom_h_predictor_8x8_sse2
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_4x4 aom_hadamard_4x4_sse2
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_8x8 aom_hadamard_8x8_sse2
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+#define aom_hadamard_lp_8x8 aom_hadamard_lp_8x8_sse2
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8_dual)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_sse2
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x128 aom_highbd_10_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x64 aom_highbd_10_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x64 aom_highbd_10_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x128 aom_highbd_10_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x16 aom_highbd_10_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x128 aom_highbd_10_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x64 aom_highbd_10_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x4 aom_highbd_10_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x64 aom_highbd_10_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x8 aom_highbd_10_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x128 aom_highbd_10_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x16 aom_highbd_10_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x32 aom_highbd_10_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x4 aom_highbd_10_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x64 aom_highbd_10_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x8 aom_highbd_10_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x16 aom_highbd_10_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x32 aom_highbd_10_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_sse2
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_sse2
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x128 aom_highbd_12_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x64 aom_highbd_12_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x64 aom_highbd_12_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x128 aom_highbd_12_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x16 aom_highbd_12_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x128 aom_highbd_12_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x64 aom_highbd_12_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x4 aom_highbd_12_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x64 aom_highbd_12_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x8 aom_highbd_12_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x128 aom_highbd_12_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x16 aom_highbd_12_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x32 aom_highbd_12_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x128 aom_highbd_12_sub_pixel_variance128x128_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x64 aom_highbd_12_sub_pixel_variance128x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x4 aom_highbd_12_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x64 aom_highbd_12_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x8 aom_highbd_12_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x128 aom_highbd_12_sub_pixel_variance64x128_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x16 aom_highbd_12_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x32 aom_highbd_12_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x128 aom_highbd_12_variance128x128_sse2
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x64 aom_highbd_12_variance128x64_sse2
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_sse2
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_sse2
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x64 aom_highbd_12_variance16x64_sse2
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_sse2
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_sse2
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_sse2
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_sse2
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x8 aom_highbd_12_variance32x8_sse2
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x128 aom_highbd_12_variance64x128_sse2
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x16 aom_highbd_12_variance64x16_sse2
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_sse2
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_sse2
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_sse2
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x32 aom_highbd_12_variance8x32_sse2
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_sse2
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_sse2
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_sse2
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x128 aom_highbd_8_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x64 aom_highbd_8_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x16 aom_highbd_8_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x32 aom_highbd_8_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x4 aom_highbd_8_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x64 aom_highbd_8_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x8 aom_highbd_8_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x16 aom_highbd_8_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x32 aom_highbd_8_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x64 aom_highbd_8_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x8 aom_highbd_8_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x16 aom_highbd_8_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x4 aom_highbd_8_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x8 aom_highbd_8_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x128 aom_highbd_8_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x16 aom_highbd_8_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x32 aom_highbd_8_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x64 aom_highbd_8_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x16 aom_highbd_8_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x32 aom_highbd_8_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x4 aom_highbd_8_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x8 aom_highbd_8_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x128 aom_highbd_8_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x64 aom_highbd_8_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x4 aom_highbd_8_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x64 aom_highbd_8_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x8 aom_highbd_8_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x128 aom_highbd_8_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x16 aom_highbd_8_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x32 aom_highbd_8_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x128 aom_highbd_8_sub_pixel_variance128x128_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x64 aom_highbd_8_sub_pixel_variance128x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x4 aom_highbd_8_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x64 aom_highbd_8_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x8 aom_highbd_8_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x128 aom_highbd_8_sub_pixel_variance64x128_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x16 aom_highbd_8_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x32 aom_highbd_8_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x128 aom_highbd_8_variance128x128_sse2
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x64 aom_highbd_8_variance128x64_sse2
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_sse2
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_sse2
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x64 aom_highbd_8_variance16x64_sse2
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_sse2
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_sse2
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_sse2
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_sse2
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x8 aom_highbd_8_variance32x8_sse2
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x128 aom_highbd_8_variance64x128_sse2
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x16 aom_highbd_8_variance64x16_sse2
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_sse2
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_sse2
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_sse2
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x32 aom_highbd_8_variance8x32_sse2
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_sse2
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+#define aom_highbd_avg_4x4 aom_highbd_avg_4x4_c
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+#define aom_highbd_avg_8x8 aom_highbd_avg_8x8_c
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+#define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_highbd_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_sse2
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_sse2
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x4 aom_highbd_dc_128_predictor_16x4_c
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x64 aom_highbd_dc_128_predictor_16x64_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_sse2
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x64 aom_highbd_dc_128_predictor_32x64_c
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x8 aom_highbd_dc_128_predictor_32x8_c
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x16 aom_highbd_dc_128_predictor_4x16_c
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_sse2
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_sse2
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x16 aom_highbd_dc_128_predictor_64x16_c
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x32 aom_highbd_dc_128_predictor_64x32_c
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x64 aom_highbd_dc_128_predictor_64x64_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_sse2
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x32 aom_highbd_dc_128_predictor_8x32_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_sse2
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_sse2
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_sse2
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_sse2
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x4 aom_highbd_dc_left_predictor_16x4_c
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x64 aom_highbd_dc_left_predictor_16x64_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_sse2
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x64 aom_highbd_dc_left_predictor_32x64_c
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x8 aom_highbd_dc_left_predictor_32x8_c
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x16 aom_highbd_dc_left_predictor_4x16_c
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_sse2
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_sse2
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x16 aom_highbd_dc_left_predictor_64x16_c
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x32 aom_highbd_dc_left_predictor_64x32_c
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x64 aom_highbd_dc_left_predictor_64x64_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_sse2
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x32 aom_highbd_dc_left_predictor_8x32_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_sse2
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_sse2
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_sse2
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_sse2
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x4 aom_highbd_dc_predictor_16x4_c
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x64 aom_highbd_dc_predictor_16x64_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_sse2
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x64 aom_highbd_dc_predictor_32x64_c
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x8 aom_highbd_dc_predictor_32x8_c
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x16 aom_highbd_dc_predictor_4x16_c
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_sse2
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_sse2
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x16 aom_highbd_dc_predictor_64x16_c
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x32 aom_highbd_dc_predictor_64x32_c
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x64 aom_highbd_dc_predictor_64x64_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_sse2
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x32 aom_highbd_dc_predictor_8x32_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_sse2
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_sse2
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_sse2
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_sse2
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x4 aom_highbd_dc_top_predictor_16x4_c
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x64 aom_highbd_dc_top_predictor_16x64_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_sse2
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x64 aom_highbd_dc_top_predictor_32x64_c
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x8 aom_highbd_dc_top_predictor_32x8_c
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x16 aom_highbd_dc_top_predictor_4x16_c
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_sse2
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_sse2
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x16 aom_highbd_dc_top_predictor_64x16_c
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x32 aom_highbd_dc_top_predictor_64x32_c
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x64 aom_highbd_dc_top_predictor_64x64_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_sse2
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x32 aom_highbd_dc_top_predictor_8x32_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_sse2
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_sse2
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_highbd_dist_wtd_comp_avg_pred_sse2(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_highbd_dist_wtd_comp_avg_pred aom_highbd_dist_wtd_comp_avg_pred_sse2
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_sse2
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_sse2
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x4 aom_highbd_h_predictor_16x4_c
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x64 aom_highbd_h_predictor_16x64_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_sse2
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x64 aom_highbd_h_predictor_32x64_c
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x8 aom_highbd_h_predictor_32x8_c
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x16 aom_highbd_h_predictor_4x16_c
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_sse2
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_sse2
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x16 aom_highbd_h_predictor_64x16_c
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x32 aom_highbd_h_predictor_64x32_c
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x64 aom_highbd_h_predictor_64x64_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_sse2
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x32 aom_highbd_h_predictor_8x32_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_sse2
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_sse2
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_14 aom_highbd_lpf_horizontal_14_sse2
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_sse2
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_6 aom_highbd_lpf_horizontal_6_sse2
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_6_dual aom_highbd_lpf_horizontal_6_dual_sse2
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_8 aom_highbd_lpf_horizontal_8_sse2
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_14 aom_highbd_lpf_vertical_14_sse2
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_4 aom_highbd_lpf_vertical_4_sse2
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_6 aom_highbd_lpf_vertical_6_sse2
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_6_dual aom_highbd_lpf_vertical_6_dual_sse2
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_8 aom_highbd_lpf_vertical_8_sse2
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_64x64_adaptive aom_highbd_quantize_b_64x64_adaptive_sse2
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x16 aom_highbd_sad4x16_sse2
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x16_avg aom_highbd_sad4x16_avg_sse2
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x3d aom_highbd_sad4x16x3d_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x4d aom_highbd_sad4x16x4d_sse2
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x4 aom_highbd_sad4x4_sse2
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x4_avg aom_highbd_sad4x4_avg_sse2
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x3d aom_highbd_sad4x4x3d_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x4d aom_highbd_sad4x4x4d_sse2
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x8 aom_highbd_sad4x8_sse2
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x8_avg aom_highbd_sad4x8_avg_sse2
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x3d aom_highbd_sad4x8x3d_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x4d aom_highbd_sad4x8x4d_sse2
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x16 aom_highbd_sad8x16_sse2
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x16_avg aom_highbd_sad8x16_avg_sse2
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x3d aom_highbd_sad8x16x3d_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x4d aom_highbd_sad8x16x4d_sse2
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x32 aom_highbd_sad8x32_sse2
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x32_avg aom_highbd_sad8x32_avg_sse2
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x3d aom_highbd_sad8x32x3d_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x4d aom_highbd_sad8x32x4d_sse2
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x4 aom_highbd_sad8x4_sse2
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x4_avg aom_highbd_sad8x4_avg_sse2
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x3d aom_highbd_sad8x4x3d_c
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x4d aom_highbd_sad8x4x4d_sse2
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x8 aom_highbd_sad8x8_sse2
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x8_avg aom_highbd_sad8x8_avg_sse2
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x3d aom_highbd_sad8x8x3d_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x4d aom_highbd_sad8x8x4d_sse2
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x4 aom_highbd_sad_skip_16x4_c
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x4x4d aom_highbd_sad_skip_16x4x4d_c
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x16 aom_highbd_sad_skip_4x16_sse2
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x16x4d aom_highbd_sad_skip_4x16x4d_sse2
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x4 aom_highbd_sad_skip_4x4_c
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x4x4d aom_highbd_sad_skip_4x4x4d_c
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x8 aom_highbd_sad_skip_4x8_sse2
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x8x4d aom_highbd_sad_skip_4x8x4d_sse2
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x16 aom_highbd_sad_skip_8x16_sse2
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x16x4d aom_highbd_sad_skip_8x16x4d_sse2
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x32 aom_highbd_sad_skip_8x32_sse2
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x32x4d aom_highbd_sad_skip_8x32x4d_sse2
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x4 aom_highbd_sad_skip_8x4_c
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x4x4d aom_highbd_sad_skip_8x4x4d_c
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x8 aom_highbd_sad_skip_8x8_sse2
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x8x4d aom_highbd_sad_skip_8x8x4d_sse2
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_highbd_sse)(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define aom_highbd_subtract_block aom_highbd_subtract_block_sse2
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_sse2
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_sse2
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_sse2
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_sse2
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_sse2
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_sse2
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_sse2
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_sse2
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output);
+#define aom_ifft4x4_float aom_ifft4x4_float_sse2
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_avx2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_col)(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_avx2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_14 aom_lpf_horizontal_14_sse2
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_minmax_8x8 aom_minmax_8x8_sse2
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_sse2
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_sse2
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_sse2
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_16xh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_sse2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_avx2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit_highbd)(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32_adaptive aom_quantize_b_32x32_adaptive_sse2
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64_adaptive aom_quantize_b_64x64_adaptive_sse2
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x16 aom_sad16x16_sse2
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_sse2
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_sse2
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_sse2
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x4 aom_sad16x4_sse2
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x4_avg aom_sad16x4_avg_sse2
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x3d aom_sad16x4x3d_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x64 aom_sad16x64_sse2
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x64_avg aom_sad16x64_avg_sse2
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x8 aom_sad16x8_sse2
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_sse2
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x8 aom_sad32x8_sse2
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x8_avg aom_sad32x8_avg_sse2
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x16 aom_sad4x16_sse2
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x16_avg aom_sad4x16_avg_sse2
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x3d aom_sad4x16x3d_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x4d aom_sad4x16x4d_sse2
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x4 aom_sad4x4_sse2
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_sse2
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x3d aom_sad4x4x3d_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x4d aom_sad4x4x4d_sse2
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_sse2
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_sse2
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x3d aom_sad4x8x3d_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x4d aom_sad4x8x4d_sse2
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x16 aom_sad64x16_sse2
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x16_avg aom_sad64x16_avg_sse2
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x16 aom_sad8x16_sse2
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_sse2
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x3d aom_sad8x16x3d_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x4d aom_sad8x16x4d_sse2
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x32 aom_sad8x32_sse2
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x32_avg aom_sad8x32_avg_sse2
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x3d aom_sad8x32x3d_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x4d aom_sad8x32x4d_sse2
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_sse2
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_sse2
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x3d aom_sad8x4x3d_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x4d aom_sad8x4x4d_sse2
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x8 aom_sad8x8_sse2
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_sse2
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x3d aom_sad8x8x3d_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x4d aom_sad8x8x4d_sse2
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x16 aom_sad_skip_16x16_sse2
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x32 aom_sad_skip_16x32_sse2
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x4 aom_sad_skip_16x4_c
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x64 aom_sad_skip_16x64_sse2
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x8 aom_sad_skip_16x8_sse2
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x8 aom_sad_skip_32x8_sse2
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x16 aom_sad_skip_4x16_sse2
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x16x4d aom_sad_skip_4x16x4d_sse2
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x4 aom_sad_skip_4x4_c
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x4x4d aom_sad_skip_4x4x4d_c
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x8 aom_sad_skip_4x8_sse2
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x8x4d aom_sad_skip_4x8x4d_sse2
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x16 aom_sad_skip_64x16_sse2
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x16 aom_sad_skip_8x16_sse2
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x16x4d aom_sad_skip_8x16x4d_sse2
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x32 aom_sad_skip_8x32_sse2
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x32x4d aom_sad_skip_8x32x4d_sse2
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x4 aom_sad_skip_8x4_c
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x4x4d aom_sad_skip_8x4x4d_c
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x8 aom_sad_skip_8x8_sse2
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x8x4d aom_sad_skip_8x8x4d_sse2
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+int aom_satd_sse2(const tran_low_t *coeff, int length);
+int aom_satd_avx2(const tran_low_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const tran_low_t *coeff, int length);
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+int aom_satd_lp_sse2(const int16_t *coeff, int length);
+int aom_satd_lp_avx2(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd_lp)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_sse)(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+void aom_ssim_parms_8x8_sse2(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_sse2
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t N);
+#define aom_sum_squares_i16 aom_sum_squares_i16_sse2
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width, int height, int *sum);
+RTCD_EXTERN uint64_t (*aom_sum_sse_2d_i16)(const int16_t *src, int src_stride, int width, int height, int *sum);
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x16 aom_v_predictor_16x16_sse2
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x4 aom_v_predictor_4x4_sse2
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_sse2
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x8 aom_v_predictor_8x8_sse2
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u16)(uint8_t *src, int src_stride, int width, int height);
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u8)(uint8_t *src, int src_stride, int width, int height);
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x16 aom_variance4x16_sse2
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x4 aom_variance4x4_sse2
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x8 aom_variance4x8_sse2
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x16 aom_variance8x16_sse2
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x32 aom_variance8x32_sse2
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x4 aom_variance8x4_sse2
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x8 aom_variance8x8_sse2
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+
+void aom_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_avg_8x8_quad = aom_avg_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_avg_8x8_quad = aom_avg_8x8_quad_avx2;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
+    aom_comp_avg_pred = aom_comp_avg_pred_c;
+    if (flags & HAS_AVX2) aom_comp_avg_pred = aom_comp_avg_pred_avx2;
+    aom_comp_mask_pred = aom_comp_mask_pred_c;
+    if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
+    if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_flow_at_point = aom_compute_flow_at_point_c;
+    if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_horiz = aom_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_horiz = aom_convolve8_horiz_avx2;
+    aom_convolve8_vert = aom_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
+    aom_convolve_copy = aom_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_avx2;
+    aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
+    aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
+    aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_avx2;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_avx2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_avx2;
+    aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
+    aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
+    aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_avx2;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_avx2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_avx2;
+    aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
+    aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
+    aom_dc_predictor_64x64 = aom_dc_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_avx2;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_avx2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_avx2;
+    aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
+    aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
+    aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_avx2;
+    aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3;
+    aom_fft16x16_float = aom_fft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_fft16x16_float = aom_fft16x16_float_avx2;
+    aom_fft32x32_float = aom_fft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_fft32x32_float = aom_fft32x32_float_avx2;
+    aom_fft8x8_float = aom_fft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_fft8x8_float = aom_fft8x8_float_avx2;
+    aom_get_blk_sse_sum = aom_get_blk_sse_sum_sse2;
+    if (flags & HAS_AVX2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_avx2;
+    aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_avx2;
+    aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_avx2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_avx2;
+    aom_hadamard_16x16 = aom_hadamard_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_16x16 = aom_hadamard_16x16_avx2;
+    aom_hadamard_32x32 = aom_hadamard_32x32_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_32x32 = aom_hadamard_32x32_avx2;
+    aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_avx2;
+    aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_avx2;
+    aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
+    aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_sse4_1;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_sse4_1;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_sse4_1;
+    aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_sse4_1;
+    aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_sse4_1;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_sse4_1;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_sse4_1;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_sse4_1;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_sse4_1;
+    aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_sse4_1;
+    aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_sse4_1;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_sse4_1;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_sse4_1;
+    aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_sse4_1;
+    aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_sse4_1;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_sse4_1;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_sse4_1;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_sse4_1;
+    aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_sse4_1;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_sse4_1;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_avx2;
+    aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_avx2;
+    aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_avx2;
+    aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_avx2;
+    aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_avx2;
+    aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_avx2;
+    aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_avx2;
+    aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_avx2;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_avx2;
+    aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_avx2;
+    aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_avx2;
+    aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_avx2;
+    aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_avx2;
+    aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_avx2;
+    aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_avx2;
+    aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_avx2;
+    aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_avx2;
+    aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_avx2;
+    aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_avx2;
+    aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_avx2;
+    aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_avx2;
+    aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_avx2;
+    aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_avx2;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_sse4_1;
+    aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_avx2;
+    aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_avx2;
+    aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_avx2;
+    aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_avx2;
+    aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_avx2;
+    aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_avx2;
+    aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_avx2;
+    aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_sse4_1;
+    aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_sse4_1;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_sse4_1;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_sse4_1;
+    aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_sse4_1;
+    aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_sse4_1;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_sse4_1;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_sse4_1;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_sse4_1;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_sse4_1;
+    aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_sse4_1;
+    aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_sse4_1;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_sse4_1;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_sse4_1;
+    aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_sse4_1;
+    aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_sse4_1;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_sse4_1;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_sse4_1;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_sse4_1;
+    aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_sse4_1;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_sse4_1;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_sse4_1;
+    aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_sse4_1;
+    aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_sse4_1;
+    aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_sse4_1;
+    aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_sse4_1;
+    aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_sse4_1;
+    aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_sse4_1;
+    aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_sse4_1;
+    aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_sse4_1;
+    aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_sse4_1;
+    aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_sse4_1;
+    aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_sse4_1;
+    aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_sse4_1;
+    aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_sse4_1;
+    aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_sse4_1;
+    aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_sse4_1;
+    aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_sse4_1;
+    aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_sse4_1;
+    aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_sse4_1;
+    aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_sse4_1;
+    aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_sse4_1;
+    aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_sse4_1;
+    aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_sse4_1;
+    aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_avx2;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_sse4_1;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_sse4_1;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
+    aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_sse2;
+    if (flags & HAS_AVX2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_avx2;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_avx2;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_avx2;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_avx2;
+    aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_avx2;
+    aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_avx2;
+    aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_avx2;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_avx2;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_avx2;
+    aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_avx2;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_avx2;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
+    aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_avx2;
+    aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_avx2;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_avx2;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_avx2;
+    aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_avx2;
+    aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_avx2;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_avx2;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_avx2;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_avx2;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_avx2;
+    aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_avx2;
+    aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_avx2;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_avx2;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_avx2;
+    aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_avx2;
+    aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_avx2;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_avx2;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_avx2;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_avx2;
+    aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_avx2;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_avx2;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_avx2;
+    aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_avx2;
+    aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_avx2;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_avx2;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_avx2;
+    aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_avx2;
+    aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_avx2;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_avx2;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_avx2;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_avx2;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_avx2;
+    aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_avx2;
+    aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_avx2;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_avx2;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_avx2;
+    aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_avx2;
+    aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_avx2;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_avx2;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_avx2;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_avx2;
+    aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_avx2;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_avx2;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_avx2;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b = aom_highbd_quantize_b_avx2;
+    aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_avx2;
+    aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_avx2;
+    aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_avx2;
+    aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_avx2;
+    aom_highbd_sad128x128 = aom_highbd_sad128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128 = aom_highbd_sad128x128_avx2;
+    aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_avx2;
+    aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_avx2;
+    aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_avx2;
+    aom_highbd_sad128x64 = aom_highbd_sad128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64 = aom_highbd_sad128x64_avx2;
+    aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_avx2;
+    aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_avx2;
+    aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_avx2;
+    aom_highbd_sad16x16 = aom_highbd_sad16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16 = aom_highbd_sad16x16_avx2;
+    aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_avx2;
+    aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_avx2;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_avx2;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32 = aom_highbd_sad16x32_avx2;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_avx2;
+    aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_avx2;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_avx2;
+    aom_highbd_sad16x4 = aom_highbd_sad16x4_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4 = aom_highbd_sad16x4_avx2;
+    aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_avx2;
+    aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_avx2;
+    aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_avx2;
+    aom_highbd_sad16x64 = aom_highbd_sad16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64 = aom_highbd_sad16x64_avx2;
+    aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_avx2;
+    aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_avx2;
+    aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_avx2;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8 = aom_highbd_sad16x8_avx2;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_avx2;
+    aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_avx2;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_avx2;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16 = aom_highbd_sad32x16_avx2;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_avx2;
+    aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_avx2;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_avx2;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32 = aom_highbd_sad32x32_avx2;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_avx2;
+    aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_avx2;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_avx2;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64 = aom_highbd_sad32x64_avx2;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_avx2;
+    aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_avx2;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_avx2;
+    aom_highbd_sad32x8 = aom_highbd_sad32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8 = aom_highbd_sad32x8_avx2;
+    aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_avx2;
+    aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_avx2;
+    aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_avx2;
+    aom_highbd_sad64x128 = aom_highbd_sad64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128 = aom_highbd_sad64x128_avx2;
+    aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_avx2;
+    aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_avx2;
+    aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_avx2;
+    aom_highbd_sad64x16 = aom_highbd_sad64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16 = aom_highbd_sad64x16_avx2;
+    aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_avx2;
+    aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_avx2;
+    aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_avx2;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32 = aom_highbd_sad64x32_avx2;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_avx2;
+    aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_avx2;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_avx2;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64 = aom_highbd_sad64x64_avx2;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_avx2;
+    aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_avx2;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_avx2;
+    aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_avx2;
+    aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_avx2;
+    aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_avx2;
+    aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_avx2;
+    aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_avx2;
+    aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_avx2;
+    aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_avx2;
+    aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_avx2;
+    aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_avx2;
+    aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_avx2;
+    aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_avx2;
+    aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_avx2;
+    aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_avx2;
+    aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_avx2;
+    aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_avx2;
+    aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_avx2;
+    aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_avx2;
+    aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_avx2;
+    aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_avx2;
+    aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_avx2;
+    aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_avx2;
+    aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_avx2;
+    aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_avx2;
+    aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_avx2;
+    aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_avx2;
+    aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_avx2;
+    aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_avx2;
+    aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_avx2;
+    aom_highbd_sse = aom_highbd_sse_c;
+    if (flags & HAS_SSE4_1) aom_highbd_sse = aom_highbd_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_sse = aom_highbd_sse_avx2;
+    aom_ifft16x16_float = aom_ifft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft16x16_float = aom_ifft16x16_float_avx2;
+    aom_ifft32x32_float = aom_ifft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft32x32_float = aom_ifft32x32_float_avx2;
+    aom_ifft8x8_float = aom_ifft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft8x8_float = aom_ifft8x8_float_avx2;
+    aom_int_pro_col = aom_int_pro_col_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_col = aom_int_pro_col_avx2;
+    aom_int_pro_row = aom_int_pro_row_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_row = aom_int_pro_row_avx2;
+    aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2;
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2;
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2;
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_avx2;
+    aom_masked_sad128x128 = aom_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128 = aom_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x128 = aom_masked_sad128x128_avx2;
+    aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_ssse3;
+    aom_masked_sad128x64 = aom_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64 = aom_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x64 = aom_masked_sad128x64_avx2;
+    aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_ssse3;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x16 = aom_masked_sad16x16_avx2;
+    aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_ssse3;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32 = aom_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x32 = aom_masked_sad16x32_avx2;
+    aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_ssse3;
+    aom_masked_sad16x4 = aom_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4 = aom_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x4 = aom_masked_sad16x4_avx2;
+    aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_ssse3;
+    aom_masked_sad16x64 = aom_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64 = aom_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x64 = aom_masked_sad16x64_avx2;
+    aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_ssse3;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8 = aom_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x8 = aom_masked_sad16x8_avx2;
+    aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_ssse3;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16 = aom_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x16 = aom_masked_sad32x16_avx2;
+    aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_ssse3;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32 = aom_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x32 = aom_masked_sad32x32_avx2;
+    aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_ssse3;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64 = aom_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x64 = aom_masked_sad32x64_avx2;
+    aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_ssse3;
+    aom_masked_sad32x8 = aom_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8 = aom_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x8 = aom_masked_sad32x8_avx2;
+    aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_ssse3;
+    aom_masked_sad4x16 = aom_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16 = aom_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x16 = aom_masked_sad4x16_avx2;
+    aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_ssse3;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4 = aom_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x4 = aom_masked_sad4x4_avx2;
+    aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_ssse3;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8 = aom_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x8 = aom_masked_sad4x8_avx2;
+    aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_ssse3;
+    aom_masked_sad64x128 = aom_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128 = aom_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x128 = aom_masked_sad64x128_avx2;
+    aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_ssse3;
+    aom_masked_sad64x16 = aom_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16 = aom_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x16 = aom_masked_sad64x16_avx2;
+    aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_ssse3;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32 = aom_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x32 = aom_masked_sad64x32_avx2;
+    aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_ssse3;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64 = aom_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x64 = aom_masked_sad64x64_avx2;
+    aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_ssse3;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16 = aom_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x16 = aom_masked_sad8x16_avx2;
+    aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_ssse3;
+    aom_masked_sad8x32 = aom_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32 = aom_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x32 = aom_masked_sad8x32_avx2;
+    aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_ssse3;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4 = aom_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x4 = aom_masked_sad8x4_avx2;
+    aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_ssse3;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8 = aom_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x8 = aom_masked_sad8x8_avx2;
+    aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_ssse3;
+    aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_ssse3;
+    aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_ssse3;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_ssse3;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_ssse3;
+    aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_ssse3;
+    aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_ssse3;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_ssse3;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_ssse3;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_ssse3;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_ssse3;
+    aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_ssse3;
+    aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_ssse3;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_ssse3;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_ssse3;
+    aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_ssse3;
+    aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_ssse3;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_ssse3;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_ssse3;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_ssse3;
+    aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_ssse3;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_ssse3;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_ssse3;
+    aom_mse16x16 = aom_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
+    aom_mse_16xh_16bit = aom_mse_16xh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_avx2;
+    aom_mse_wxh_16bit = aom_mse_wxh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_avx2;
+    aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_avx2;
+    aom_obmc_sad128x128 = aom_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x128 = aom_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x128 = aom_obmc_sad128x128_avx2;
+    aom_obmc_sad128x64 = aom_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x64 = aom_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x64 = aom_obmc_sad128x64_avx2;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x16 = aom_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x16 = aom_obmc_sad16x16_avx2;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x32 = aom_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x32 = aom_obmc_sad16x32_avx2;
+    aom_obmc_sad16x4 = aom_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x4 = aom_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x4 = aom_obmc_sad16x4_avx2;
+    aom_obmc_sad16x64 = aom_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x64 = aom_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x64 = aom_obmc_sad16x64_avx2;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x8 = aom_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x8 = aom_obmc_sad16x8_avx2;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x16 = aom_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x16 = aom_obmc_sad32x16_avx2;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x32 = aom_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x32 = aom_obmc_sad32x32_avx2;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x64 = aom_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x64 = aom_obmc_sad32x64_avx2;
+    aom_obmc_sad32x8 = aom_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x8 = aom_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x8 = aom_obmc_sad32x8_avx2;
+    aom_obmc_sad4x16 = aom_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x16 = aom_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x16 = aom_obmc_sad4x16_avx2;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x4 = aom_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x4 = aom_obmc_sad4x4_avx2;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x8 = aom_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x8 = aom_obmc_sad4x8_avx2;
+    aom_obmc_sad64x128 = aom_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x128 = aom_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x128 = aom_obmc_sad64x128_avx2;
+    aom_obmc_sad64x16 = aom_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x16 = aom_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x16 = aom_obmc_sad64x16_avx2;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x32 = aom_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x32 = aom_obmc_sad64x32_avx2;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x64 = aom_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x64 = aom_obmc_sad64x64_avx2;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x16 = aom_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x16 = aom_obmc_sad8x16_avx2;
+    aom_obmc_sad8x32 = aom_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x32 = aom_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x32 = aom_obmc_sad8x32_avx2;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x4 = aom_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x4 = aom_obmc_sad8x4_avx2;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x8 = aom_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x8 = aom_obmc_sad8x8_avx2;
+    aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_sse4_1;
+    aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_sse4_1;
+    aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_sse4_1;
+    aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_sse4_1;
+    aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_sse4_1;
+    aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_sse4_1;
+    aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_sse4_1;
+    aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_sse4_1;
+    aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_sse4_1;
+    aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_sse4_1;
+    aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_sse4_1;
+    aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_sse4_1;
+    aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_sse4_1;
+    aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_sse4_1;
+    aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_sse4_1;
+    aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_sse4_1;
+    aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_sse4_1;
+    aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_sse4_1;
+    aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_sse4_1;
+    aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_sse4_1;
+    aom_obmc_variance128x128 = aom_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x128 = aom_obmc_variance128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x128 = aom_obmc_variance128x128_avx2;
+    aom_obmc_variance128x64 = aom_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x64 = aom_obmc_variance128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x64 = aom_obmc_variance128x64_avx2;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x16 = aom_obmc_variance16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x16 = aom_obmc_variance16x16_avx2;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x32 = aom_obmc_variance16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x32 = aom_obmc_variance16x32_avx2;
+    aom_obmc_variance16x4 = aom_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x4 = aom_obmc_variance16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x4 = aom_obmc_variance16x4_avx2;
+    aom_obmc_variance16x64 = aom_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x64 = aom_obmc_variance16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x64 = aom_obmc_variance16x64_avx2;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x8 = aom_obmc_variance16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x8 = aom_obmc_variance16x8_avx2;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x16 = aom_obmc_variance32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x16 = aom_obmc_variance32x16_avx2;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x32 = aom_obmc_variance32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x32 = aom_obmc_variance32x32_avx2;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x64 = aom_obmc_variance32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x64 = aom_obmc_variance32x64_avx2;
+    aom_obmc_variance32x8 = aom_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x8 = aom_obmc_variance32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x8 = aom_obmc_variance32x8_avx2;
+    aom_obmc_variance4x16 = aom_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x16 = aom_obmc_variance4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x16 = aom_obmc_variance4x16_avx2;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x4 = aom_obmc_variance4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x4 = aom_obmc_variance4x4_avx2;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x8 = aom_obmc_variance4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x8 = aom_obmc_variance4x8_avx2;
+    aom_obmc_variance64x128 = aom_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x128 = aom_obmc_variance64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x128 = aom_obmc_variance64x128_avx2;
+    aom_obmc_variance64x16 = aom_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x16 = aom_obmc_variance64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x16 = aom_obmc_variance64x16_avx2;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x32 = aom_obmc_variance64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x32 = aom_obmc_variance64x32_avx2;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x64 = aom_obmc_variance64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x64 = aom_obmc_variance64x64_avx2;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x16 = aom_obmc_variance8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x16 = aom_obmc_variance8x16_avx2;
+    aom_obmc_variance8x32 = aom_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x32 = aom_obmc_variance8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x32 = aom_obmc_variance8x32_avx2;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x4 = aom_obmc_variance8x4_avx2;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x8 = aom_obmc_variance8x8_avx2;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_avx2;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_avx2;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_avx2;
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_avx2;
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_avx2;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3;
+    aom_quantize_b = aom_quantize_b_sse2;
+    if (flags & HAS_SSSE3) aom_quantize_b = aom_quantize_b_ssse3;
+    if (flags & HAS_AVX) aom_quantize_b = aom_quantize_b_avx;
+    if (flags & HAS_AVX2) aom_quantize_b = aom_quantize_b_avx2;
+    aom_quantize_b_32x32 = aom_quantize_b_32x32_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_32x32 = aom_quantize_b_32x32_ssse3;
+    if (flags & HAS_AVX) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx;
+    if (flags & HAS_AVX2) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx2;
+    aom_quantize_b_64x64 = aom_quantize_b_64x64_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_64x64 = aom_quantize_b_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_quantize_b_64x64 = aom_quantize_b_64x64_avx2;
+    aom_quantize_b_adaptive = aom_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_avx2;
+    aom_sad128x128 = aom_sad128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128 = aom_sad128x128_avx2;
+    aom_sad128x128_avg = aom_sad128x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128_avg = aom_sad128x128_avg_avx2;
+    aom_sad128x128x3d = aom_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x128x3d = aom_sad128x128x3d_avx2;
+    aom_sad128x128x4d = aom_sad128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128x4d = aom_sad128x128x4d_avx2;
+    aom_sad128x64 = aom_sad128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64 = aom_sad128x64_avx2;
+    aom_sad128x64_avg = aom_sad128x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64_avg = aom_sad128x64_avg_avx2;
+    aom_sad128x64x3d = aom_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x64x3d = aom_sad128x64x3d_avx2;
+    aom_sad128x64x4d = aom_sad128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64x4d = aom_sad128x64x4d_avx2;
+    aom_sad16x16x3d = aom_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x16x3d = aom_sad16x16x3d_avx2;
+    aom_sad16x16x4d = aom_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x16x4d = aom_sad16x16x4d_avx2;
+    aom_sad16x32x3d = aom_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
+    aom_sad16x32x4d = aom_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4x4d = aom_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
+    aom_sad16x64x3d = aom_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x64x3d = aom_sad16x64x3d_avx2;
+    aom_sad16x64x4d = aom_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x64x4d = aom_sad16x64x4d_avx2;
+    aom_sad16x8x3d = aom_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x8x3d = aom_sad16x8x3d_avx2;
+    aom_sad16x8x4d = aom_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x8x4d = aom_sad16x8x4d_avx2;
+    aom_sad32x16 = aom_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16 = aom_sad32x16_avx2;
+    aom_sad32x16_avg = aom_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16_avg = aom_sad32x16_avg_avx2;
+    aom_sad32x16x3d = aom_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x16x3d = aom_sad32x16x3d_avx2;
+    aom_sad32x16x4d = aom_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16x4d = aom_sad32x16x4d_avx2;
+    aom_sad32x32 = aom_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32 = aom_sad32x32_avx2;
+    aom_sad32x32_avg = aom_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32_avg = aom_sad32x32_avg_avx2;
+    aom_sad32x32x3d = aom_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x32x3d = aom_sad32x32x3d_avx2;
+    aom_sad32x32x4d = aom_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32x4d = aom_sad32x32x4d_avx2;
+    aom_sad32x64 = aom_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64 = aom_sad32x64_avx2;
+    aom_sad32x64_avg = aom_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64_avg = aom_sad32x64_avg_avx2;
+    aom_sad32x64x3d = aom_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x64x3d = aom_sad32x64x3d_avx2;
+    aom_sad32x64x4d = aom_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64x4d = aom_sad32x64x4d_avx2;
+    aom_sad32x8x3d = aom_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x8x3d = aom_sad32x8x3d_avx2;
+    aom_sad32x8x4d = aom_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x8x4d = aom_sad32x8x4d_avx2;
+    aom_sad64x128 = aom_sad64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128 = aom_sad64x128_avx2;
+    aom_sad64x128_avg = aom_sad64x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128_avg = aom_sad64x128_avg_avx2;
+    aom_sad64x128x3d = aom_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x128x3d = aom_sad64x128x3d_avx2;
+    aom_sad64x128x4d = aom_sad64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128x4d = aom_sad64x128x4d_avx2;
+    aom_sad64x16x3d = aom_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x16x3d = aom_sad64x16x3d_avx2;
+    aom_sad64x16x4d = aom_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x16x4d = aom_sad64x16x4d_avx2;
+    aom_sad64x32 = aom_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32 = aom_sad64x32_avx2;
+    aom_sad64x32_avg = aom_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32_avg = aom_sad64x32_avg_avx2;
+    aom_sad64x32x3d = aom_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x32x3d = aom_sad64x32x3d_avx2;
+    aom_sad64x32x4d = aom_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32x4d = aom_sad64x32x4d_avx2;
+    aom_sad64x64 = aom_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64 = aom_sad64x64_avx2;
+    aom_sad64x64_avg = aom_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64_avg = aom_sad64x64_avg_avx2;
+    aom_sad64x64x3d = aom_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x64x3d = aom_sad64x64x3d_avx2;
+    aom_sad64x64x4d = aom_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64x4d = aom_sad64x64x4d_avx2;
+    aom_sad_skip_128x128 = aom_sad_skip_128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128 = aom_sad_skip_128x128_avx2;
+    aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_avx2;
+    aom_sad_skip_128x64 = aom_sad_skip_128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64 = aom_sad_skip_128x64_avx2;
+    aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_avx2;
+    aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
+    aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
+    aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_avx2;
+    aom_sad_skip_32x16 = aom_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16 = aom_sad_skip_32x16_avx2;
+    aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_avx2;
+    aom_sad_skip_32x32 = aom_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32 = aom_sad_skip_32x32_avx2;
+    aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_avx2;
+    aom_sad_skip_32x64 = aom_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64 = aom_sad_skip_32x64_avx2;
+    aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_avx2;
+    aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_avx2;
+    aom_sad_skip_64x128 = aom_sad_skip_64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128 = aom_sad_skip_64x128_avx2;
+    aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_avx2;
+    aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_avx2;
+    aom_sad_skip_64x32 = aom_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32 = aom_sad_skip_64x32_avx2;
+    aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_avx2;
+    aom_sad_skip_64x64 = aom_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64 = aom_sad_skip_64x64_avx2;
+    aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_avx2;
+    aom_satd = aom_satd_sse2;
+    if (flags & HAS_AVX2) aom_satd = aom_satd_avx2;
+    aom_satd_lp = aom_satd_lp_sse2;
+    if (flags & HAS_AVX2) aom_satd_lp = aom_satd_lp_avx2;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3;
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_ssse3;
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_ssse3;
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_ssse3;
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_ssse3;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_ssse3;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_ssse3;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_ssse3;
+    aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
+    aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
+    aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_ssse3;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_ssse3;
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_ssse3;
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_ssse3;
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_ssse3;
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_ssse3;
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_ssse3;
+    aom_sse = aom_sse_c;
+    if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_subtract_block = aom_subtract_block_sse2;
+    if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_avx2;
+    aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_avx2;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x16 = aom_v_predictor_32x16_avx2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x32 = aom_v_predictor_32x32_avx2;
+    aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
+    aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
+    aom_v_predictor_64x64 = aom_v_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x64 = aom_v_predictor_64x64_avx2;
+    aom_var_2d_u16 = aom_var_2d_u16_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u16 = aom_var_2d_u16_avx2;
+    aom_var_2d_u8 = aom_var_2d_u8_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u8 = aom_var_2d_u8_avx2;
+    aom_variance128x128 = aom_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_variance128x128 = aom_variance128x128_avx2;
+    aom_variance128x64 = aom_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_variance128x64 = aom_variance128x64_avx2;
+    aom_variance16x16 = aom_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
+    aom_variance16x32 = aom_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_variance16x32 = aom_variance16x32_avx2;
+    aom_variance16x4 = aom_variance16x4_sse2;
+    if (flags & HAS_AVX2) aom_variance16x4 = aom_variance16x4_avx2;
+    aom_variance16x64 = aom_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_variance16x64 = aom_variance16x64_avx2;
+    aom_variance16x8 = aom_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_variance16x8 = aom_variance16x8_avx2;
+    aom_variance32x16 = aom_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_variance32x16 = aom_variance32x16_avx2;
+    aom_variance32x32 = aom_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_variance32x32 = aom_variance32x32_avx2;
+    aom_variance32x64 = aom_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_variance32x64 = aom_variance32x64_avx2;
+    aom_variance32x8 = aom_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_variance32x8 = aom_variance32x8_avx2;
+    aom_variance64x128 = aom_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_variance64x128 = aom_variance64x128_avx2;
+    aom_variance64x16 = aom_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_variance64x16 = aom_variance64x16_avx2;
+    aom_variance64x32 = aom_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_variance64x32 = aom_variance64x32_avx2;
+    aom_variance64x64 = aom_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_variance64x64 = aom_variance64x64_avx2;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
+    if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
+    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
+    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/mac/x64/config/aom_scale_rtcd.h b/media/libaom/config/mac/x64/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..3b70fb47c3
--- /dev/null
+++ b/media/libaom/config/mac/x64/config/aom_scale_rtcd.h
@@ -0,0 +1,103 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/mac/x64/config/av1_rtcd.h b/media/libaom/config/mac/x64/config/av1_rtcd.h
new file mode 100644
index 0000000000..b1cdc99700
--- /dev/null
+++ b/media/libaom/config/mac/x64/config/av1_rtcd.h
@@ -0,0 +1,1217 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+void aom_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_sse2
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_dist_wtd_comp_avg_upsampled_pred_ssse3(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_sse2
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+#define aom_highbd_dist_wtd_comp_avg_upsampled_pred aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_upsampled_pred aom_highbd_upsampled_pred_sse2
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+#define aom_quantize_b_helper aom_quantize_b_helper_c
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+#define aom_upsampled_pred aom_upsampled_pred_sse2
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_sse4_1(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_avx2(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_avx2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+RTCD_EXTERN int64_t (*av1_block_error_lp)(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim1)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim2)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params_high_bd)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+void av1_cnn_convolve_no_maxpool_padding_valid_avx2(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+RTCD_EXTERN void (*av1_cnn_convolve_no_maxpool_padding_valid)(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+RTCD_EXTERN void (*av1_compute_stats)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+RTCD_EXTERN void (*av1_compute_stats_highbd)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr_intrabc av1_convolve_2d_sr_intrabc_c
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr_intrabc av1_convolve_x_sr_intrabc_c
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+double av1_estimate_noise_from_single_plane_avx2(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+RTCD_EXTERN double (*av1_estimate_noise_from_single_plane)(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void av1_fwht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*av1_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+uint32_t av1_get_crc32c_value_sse4_2(void *crc_calculator, uint8_t *p, size_t length);
+RTCD_EXTERN uint32_t (*av1_get_crc32c_value)(void *crc_calculator, uint8_t *p, size_t length);
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_sse4_1(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_avx2(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+RTCD_EXTERN void (*av1_get_horver_correlation_full)(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+void av1_get_nz_map_contexts_sse2(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+#define av1_get_nz_map_contexts av1_get_nz_map_contexts_sse2
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_highbd_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+RTCD_EXTERN int64_t (*av1_highbd_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_sse2
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_sse2
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr_intrabc av1_highbd_convolve_2d_sr_intrabc_c
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr_intrabc av1_highbd_convolve_x_sr_intrabc_c
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr_intrabc av1_highbd_convolve_y_sr_intrabc_c
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+#define av1_highbd_estimate_noise_from_single_plane av1_highbd_estimate_noise_from_single_plane_c
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+void av1_highbd_filter_intra_edge_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_highbd_filter_intra_edge)(uint16_t *p, int sz, int strength);
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_highbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_sse4_1(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+RTCD_EXTERN void (*av1_highbd_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+void av1_highbd_upsample_intra_edge_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_highbd_upsample_intra_edge)(uint16_t *p, int sz, int bd);
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_lowbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+void av1_nn_fast_softmax_16_sse3(const float *input_nodes, float *output);
+RTCD_EXTERN void (*av1_nn_fast_softmax_16)(const float *input_nodes, float *output);
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_sse3(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_avx2(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+RTCD_EXTERN void (*av1_nn_predict)(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_avx2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_lp)(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit);
+RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit);
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+RTCD_EXTERN void (*av1_txb_init_levels)(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_avx2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+RTCD_EXTERN void (*av1_wedge_compute_delta_squares)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_avx2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+RTCD_EXTERN int8_t (*av1_wedge_sign_from_residuals)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+RTCD_EXTERN uint64_t (*av1_wedge_sse_from_residuals)(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*cfl_get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*cfl_get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*cfl_get_subtract_average_fn)(TX_SIZE tx_size);
+
+void av1_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
+    av1_apply_temporal_filter = av1_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_apply_temporal_filter = av1_apply_temporal_filter_avx2;
+    av1_block_error = av1_block_error_sse2;
+    if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_block_error_lp = av1_block_error_lp_sse2;
+    if (flags & HAS_AVX2) av1_block_error_lp = av1_block_error_lp_avx2;
+    av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
+    av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
+    av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
+    if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
+    av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
+    av1_calc_indices_dim2 = av1_calc_indices_dim2_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim2 = av1_calc_indices_dim2_avx2;
+    av1_calc_proj_params = av1_calc_proj_params_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params = av1_calc_proj_params_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params = av1_calc_proj_params_avx2;
+    av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_avx2;
+    av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_c;
+    if (flags & HAS_AVX2) av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_avx2;
+    av1_compute_stats = av1_compute_stats_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats = av1_compute_stats_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats = av1_compute_stats_avx2;
+    av1_compute_stats_highbd = av1_compute_stats_highbd_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats_highbd = av1_compute_stats_highbd_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats_highbd = av1_compute_stats_highbd_avx2;
+    av1_convolve_2d_scale = av1_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
+    av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
+    av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
+    av1_convolve_x_sr = av1_convolve_x_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
+    av1_convolve_y_sr = av1_convolve_y_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
+    av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_avx2;
+    av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_avx2;
+    av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2;
+    av1_dr_prediction_z1 = av1_dr_prediction_z1_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2;
+    av1_dr_prediction_z2 = av1_dr_prediction_z2_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2;
+    av1_dr_prediction_z3 = av1_dr_prediction_z3_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2;
+    av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_c;
+    if (flags & HAS_AVX2) av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_avx2;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
+    av1_filter_intra_predictor = av1_filter_intra_predictor_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_avx2;
+    av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_sse4_1;
+    av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_sse4_1;
+    av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_sse4_1;
+    av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_avx2;
+    av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_sse4_1;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_avx2;
+    av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_sse4_1;
+    av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_sse4_1;
+    av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_sse4_1;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
+    av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_sse4_1;
+    av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_sse4_1;
+    av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_sse4_1;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_avx2;
+    av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_avx2;
+    av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_sse4_1;
+    av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_sse4_1;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_avx2;
+    av1_fwht4x4 = av1_fwht4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwht4x4 = av1_fwht4x4_sse4_1;
+    av1_get_crc32c_value = av1_get_crc32c_value_c;
+    if (flags & HAS_SSE4_2) av1_get_crc32c_value = av1_get_crc32c_value_sse4_2;
+    av1_get_horver_correlation_full = av1_get_horver_correlation_full_c;
+    if (flags & HAS_SSE4_1) av1_get_horver_correlation_full = av1_get_horver_correlation_full_sse4_1;
+    if (flags & HAS_AVX2) av1_get_horver_correlation_full = av1_get_horver_correlation_full_avx2;
+    av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_avx2;
+    av1_highbd_block_error = av1_highbd_block_error_sse2;
+    if (flags & HAS_AVX2) av1_highbd_block_error = av1_highbd_block_error_avx2;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
+    av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
+    av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
+    av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
+    av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_avx2;
+    av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_avx2;
+    av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_avx2;
+    av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_avx2;
+    av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_avx2;
+    av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_avx2;
+    av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_avx2;
+    av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_sse4_1;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_sse4_1;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_sse4_1;
+    av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
+    av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1;
+    av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_avx2;
+    av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
+    if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
+    av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_sse4_1;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2;
+    av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
+    if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
+    av1_inv_txfm_add = av1_inv_txfm_add_c;
+    if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
+    if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
+    av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_avx2;
+    av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_c;
+    if (flags & HAS_SSE3) av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_sse3;
+    av1_nn_predict = av1_nn_predict_c;
+    if (flags & HAS_SSE3) av1_nn_predict = av1_nn_predict_sse3;
+    if (flags & HAS_AVX2) av1_nn_predict = av1_nn_predict_avx2;
+    av1_quantize_fp = av1_quantize_fp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
+    av1_quantize_fp_64x64 = av1_quantize_fp_64x64_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_64x64 = av1_quantize_fp_64x64_avx2;
+    av1_quantize_lp = av1_quantize_lp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_lp = av1_quantize_lp_avx2;
+    av1_resize_and_extend_frame = av1_resize_and_extend_frame_c;
+    if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
+    av1_txb_init_levels = av1_txb_init_levels_c;
+    if (flags & HAS_SSE4_1) av1_txb_init_levels = av1_txb_init_levels_sse4_1;
+    if (flags & HAS_AVX2) av1_txb_init_levels = av1_txb_init_levels_avx2;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_warp_affine = av1_warp_affine_avx2;
+    av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
+    if (flags & HAS_AVX2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
+    av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_avx2;
+    av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
+    av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
+    if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
+    cdef_filter_16_0 = cdef_filter_16_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
+    cdef_filter_16_1 = cdef_filter_16_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
+    cdef_filter_16_2 = cdef_filter_16_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
+    cdef_filter_16_3 = cdef_filter_16_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
+    cdef_filter_8_0 = cdef_filter_8_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
+    cdef_filter_8_1 = cdef_filter_8_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
+    cdef_filter_8_2 = cdef_filter_8_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
+    cdef_filter_8_3 = cdef_filter_8_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
+    cdef_find_dir = cdef_find_dir_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
+    cdef_find_dir_dual = cdef_find_dir_dual_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
+    cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
+    cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
+    cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
+    cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
+    cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
+    cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
+    cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_avx2;
+    cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_avx2;
+    cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/win/ia32/config/aom_config.asm b/media/libaom/config/win/ia32/config/aom_config.asm
new file mode 100644
index 0000000000..af78328283
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_config.asm
@@ -0,0 +1,92 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+AOM_ARCH_AARCH64 equ 0
+AOM_ARCH_ARM equ 0
+AOM_ARCH_PPC equ 0
+AOM_ARCH_X86 equ 1
+AOM_ARCH_X86_64 equ 0
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_HIGHBITDEPTH equ 1
+CONFIG_AV1_TEMPORAL_DENOISING equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITRATE_ACCURACY equ 0
+CONFIG_BITRATE_ACCURACY_BL equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_COMPONENT_TIMING equ 0
+CONFIG_COLLECT_PARTITION_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_CWG_C013 equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_EXCLUDE_SIMD_MISMATCH equ 0
+CONFIG_FPMT_TEST equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBVMAF_PSNR_PEAK equ 1
+CONFIG_LIBYUV equ 0
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NN_V2 equ 0
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OPTICAL_FLOW_API equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_OUTPUT_FRAME_SIZE equ 0
+CONFIG_PARTITION_SEARCH_ORDER equ 0
+CONFIG_PIC equ 0
+CONFIG_RATECTRL_LOG equ 0
+CONFIG_RD_COMMAND equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_RT_ML_PARTITIONING equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SALIENCY_MAP equ 0
+CONFIG_SHARED equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_SPEED_STATS equ 0
+CONFIG_TFLITE equ 0
+CONFIG_THREE_PASS equ 0
+CONFIG_TUNE_BUTTERAUGLI equ 0
+CONFIG_TUNE_VMAF equ 0
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+FORCE_HIGHBITDEPTH_DECODING equ 0
+HAVE_ARM_CRC32 equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_FEXCEPT equ 1
+HAVE_MMX equ 1
+HAVE_NEON equ 0
+HAVE_NEON_DOTPROD equ 0
+HAVE_NEON_I8MM equ 0
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 1
+HAVE_SSE4_1 equ 1
+HAVE_SSE4_2 equ 1
+HAVE_SSSE3 equ 1
+HAVE_SVE equ 0
+HAVE_VSX equ 0
+HAVE_WXWIDGETS equ 0
+STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/win/ia32/config/aom_config.h b/media/libaom/config/win/ia32/config/aom_config.h
new file mode 100644
index 0000000000..dba805b1b6
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 0
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 1
+#define AOM_ARCH_X86_64 0
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 0
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 1
+#define HAVE_NEON 0
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_SSE4_2 1
+#define HAVE_SSSE3 1
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..a19adf5f61
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h
@@ -0,0 +1,9185 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_4x4)(const uint8_t *, int p);
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_8x8)(const uint8_t *, int p);
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_avx2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_dist_wtd_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN unsigned int (*aom_dist_wtd_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+void aom_fdct4x4_lp_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4_lp)(const int16_t *input, int16_t *output, int stride);
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+void aom_fft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_fft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+void aom_fft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_fft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+void aom_fft4x4_float_sse2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft4x4_float)(const float *input, float *temp, float *output);
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+void aom_fft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_fft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_sse2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_avx2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+RTCD_EXTERN void (*aom_get_blk_sse_sum)(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+RTCD_EXTERN unsigned int (*aom_get_mb_ss)(const int16_t *);
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+RTCD_EXTERN void (*aom_get_var_sse_sum_16x16_dual)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+RTCD_EXTERN void (*aom_get_var_sse_sum_8x8_quad)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_4x4)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8_dual)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x128 aom_highbd_10_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x64 aom_highbd_10_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x64 aom_highbd_10_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x128 aom_highbd_10_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x16 aom_highbd_10_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x128 aom_highbd_10_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x64 aom_highbd_10_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x128 aom_highbd_10_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x128 aom_highbd_12_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x64 aom_highbd_12_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x64 aom_highbd_12_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x128 aom_highbd_12_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x16 aom_highbd_12_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x128 aom_highbd_12_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x64 aom_highbd_12_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x128 aom_highbd_12_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x128 aom_highbd_8_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x64 aom_highbd_8_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x16 aom_highbd_8_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x32 aom_highbd_8_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x4 aom_highbd_8_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x64 aom_highbd_8_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x8 aom_highbd_8_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x16 aom_highbd_8_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x32 aom_highbd_8_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x64 aom_highbd_8_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x8 aom_highbd_8_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x16 aom_highbd_8_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x4 aom_highbd_8_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x8 aom_highbd_8_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x128 aom_highbd_8_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x16 aom_highbd_8_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x32 aom_highbd_8_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x64 aom_highbd_8_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x16 aom_highbd_8_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x32 aom_highbd_8_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x4 aom_highbd_8_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x8 aom_highbd_8_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x128 aom_highbd_8_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x64 aom_highbd_8_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x128 aom_highbd_8_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+#define aom_highbd_avg_4x4 aom_highbd_avg_4x4_c
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+#define aom_highbd_avg_8x8 aom_highbd_avg_8x8_c
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+#define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_highbd_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x4 aom_highbd_dc_128_predictor_16x4_c
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x64 aom_highbd_dc_128_predictor_16x64_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x64 aom_highbd_dc_128_predictor_32x64_c
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x8 aom_highbd_dc_128_predictor_32x8_c
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x16 aom_highbd_dc_128_predictor_4x16_c
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x16 aom_highbd_dc_128_predictor_64x16_c
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x32 aom_highbd_dc_128_predictor_64x32_c
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x64 aom_highbd_dc_128_predictor_64x64_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x32 aom_highbd_dc_128_predictor_8x32_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_128_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x4 aom_highbd_dc_left_predictor_16x4_c
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x64 aom_highbd_dc_left_predictor_16x64_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x64 aom_highbd_dc_left_predictor_32x64_c
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x8 aom_highbd_dc_left_predictor_32x8_c
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x16 aom_highbd_dc_left_predictor_4x16_c
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x16 aom_highbd_dc_left_predictor_64x16_c
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x32 aom_highbd_dc_left_predictor_64x32_c
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x64 aom_highbd_dc_left_predictor_64x64_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x32 aom_highbd_dc_left_predictor_8x32_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_left_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x4 aom_highbd_dc_predictor_16x4_c
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x64 aom_highbd_dc_predictor_16x64_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x64 aom_highbd_dc_predictor_32x64_c
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x8 aom_highbd_dc_predictor_32x8_c
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x16 aom_highbd_dc_predictor_4x16_c
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x16 aom_highbd_dc_predictor_64x16_c
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x32 aom_highbd_dc_predictor_64x32_c
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x64 aom_highbd_dc_predictor_64x64_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x32 aom_highbd_dc_predictor_8x32_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x4 aom_highbd_dc_top_predictor_16x4_c
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x64 aom_highbd_dc_top_predictor_16x64_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x64 aom_highbd_dc_top_predictor_32x64_c
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x8 aom_highbd_dc_top_predictor_32x8_c
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x16 aom_highbd_dc_top_predictor_4x16_c
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x16 aom_highbd_dc_top_predictor_64x16_c
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x32 aom_highbd_dc_top_predictor_64x32_c
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x64 aom_highbd_dc_top_predictor_64x64_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x32 aom_highbd_dc_top_predictor_8x32_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_dc_top_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_highbd_dist_wtd_comp_avg_pred_sse2(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_highbd_dist_wtd_comp_avg_pred)(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x4 aom_highbd_h_predictor_16x4_c
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x64 aom_highbd_h_predictor_16x64_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x64 aom_highbd_h_predictor_32x64_c
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x8 aom_highbd_h_predictor_32x8_c
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x16 aom_highbd_h_predictor_4x16_c
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x16 aom_highbd_h_predictor_64x16_c
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x32 aom_highbd_h_predictor_64x32_c
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x64 aom_highbd_h_predictor_64x64_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x32 aom_highbd_h_predictor_8x32_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_h_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x3d aom_highbd_sad4x16x3d_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x3d aom_highbd_sad4x4x3d_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x3d aom_highbd_sad4x8x3d_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x3d aom_highbd_sad8x16x3d_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x3d aom_highbd_sad8x32x3d_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x3d aom_highbd_sad8x4x3d_c
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x3d aom_highbd_sad8x8x3d_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x4 aom_highbd_sad_skip_16x4_c
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x4x4d aom_highbd_sad_skip_16x4x4d_c
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x4 aom_highbd_sad_skip_4x4_c
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x4x4d aom_highbd_sad_skip_4x4x4d_c
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x4 aom_highbd_sad_skip_8x4_c
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x4x4d aom_highbd_sad_skip_8x4x4d_c
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_highbd_sse)(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_highbd_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft4x4_float)(const float *input, float *temp, float *output);
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_avx2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_col)(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_avx2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+RTCD_EXTERN void (*aom_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_16xh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_sse2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_avx2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit_highbd)(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x3d aom_sad16x4x3d_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x3d aom_sad4x16x3d_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x3d aom_sad4x4x3d_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x3d aom_sad4x8x3d_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x3d aom_sad8x16x3d_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x3d aom_sad8x32x3d_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x3d aom_sad8x4x3d_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x3d aom_sad8x8x3d_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x4 aom_sad_skip_16x4_c
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x4 aom_sad_skip_4x4_c
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x4x4d aom_sad_skip_4x4x4d_c
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x4 aom_sad_skip_8x4_c
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x4x4d aom_sad_skip_8x4x4d_c
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+int aom_satd_sse2(const tran_low_t *coeff, int length);
+int aom_satd_avx2(const tran_low_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const tran_low_t *coeff, int length);
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+int aom_satd_lp_sse2(const int16_t *coeff, int length);
+int aom_satd_lp_avx2(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd_lp)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_sse)(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_c
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t N);
+RTCD_EXTERN uint64_t (*aom_sum_squares_i16)(const int16_t *src, uint32_t N);
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width, int height, int *sum);
+RTCD_EXTERN uint64_t (*aom_sum_sse_2d_i16)(const int16_t *src, int src_stride, int width, int height, int *sum);
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u16)(uint8_t *src, int src_stride, int width, int height);
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u8)(uint8_t *src, int src_stride, int width, int height);
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+
+void aom_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_avg_4x4 = aom_avg_4x4_c;
+    if (flags & HAS_SSE2) aom_avg_4x4 = aom_avg_4x4_sse2;
+    aom_avg_8x8 = aom_avg_8x8_c;
+    if (flags & HAS_SSE2) aom_avg_8x8 = aom_avg_8x8_sse2;
+    aom_avg_8x8_quad = aom_avg_8x8_quad_c;
+    if (flags & HAS_SSE2) aom_avg_8x8_quad = aom_avg_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_avg_8x8_quad = aom_avg_8x8_quad_avx2;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
+    aom_comp_avg_pred = aom_comp_avg_pred_c;
+    if (flags & HAS_AVX2) aom_comp_avg_pred = aom_comp_avg_pred_avx2;
+    aom_comp_mask_pred = aom_comp_mask_pred_c;
+    if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
+    if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_flow_at_point = aom_compute_flow_at_point_c;
+    if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
+    if (flags & HAS_SSE2) aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_horiz = aom_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_horiz = aom_convolve8_horiz_avx2;
+    aom_convolve8_vert = aom_convolve8_vert_c;
+    if (flags & HAS_SSE2) aom_convolve8_vert = aom_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
+    aom_convolve_copy = aom_convolve_copy_c;
+    if (flags & HAS_SSE2) aom_convolve_copy = aom_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2;
+    aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_sse2;
+    aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_sse2;
+    aom_dc_128_predictor_16x4 = aom_dc_128_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x4 = aom_dc_128_predictor_16x4_sse2;
+    aom_dc_128_predictor_16x64 = aom_dc_128_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x64 = aom_dc_128_predictor_16x64_sse2;
+    aom_dc_128_predictor_16x8 = aom_dc_128_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_16x8 = aom_dc_128_predictor_16x8_sse2;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_avx2;
+    aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+    aom_dc_128_predictor_32x8 = aom_dc_128_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x8 = aom_dc_128_predictor_32x8_sse2;
+    aom_dc_128_predictor_4x16 = aom_dc_128_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x16 = aom_dc_128_predictor_4x16_sse2;
+    aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_sse2;
+    aom_dc_128_predictor_4x8 = aom_dc_128_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x8 = aom_dc_128_predictor_4x8_sse2;
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
+    aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
+    aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_avx2;
+    aom_dc_128_predictor_8x16 = aom_dc_128_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x16 = aom_dc_128_predictor_8x16_sse2;
+    aom_dc_128_predictor_8x32 = aom_dc_128_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x32 = aom_dc_128_predictor_8x32_sse2;
+    aom_dc_128_predictor_8x4 = aom_dc_128_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x4 = aom_dc_128_predictor_8x4_sse2;
+    aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_sse2;
+    aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_sse2;
+    aom_dc_left_predictor_16x32 = aom_dc_left_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x32 = aom_dc_left_predictor_16x32_sse2;
+    aom_dc_left_predictor_16x4 = aom_dc_left_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x4 = aom_dc_left_predictor_16x4_sse2;
+    aom_dc_left_predictor_16x64 = aom_dc_left_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x64 = aom_dc_left_predictor_16x64_sse2;
+    aom_dc_left_predictor_16x8 = aom_dc_left_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x8 = aom_dc_left_predictor_16x8_sse2;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_avx2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_avx2;
+    aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+    aom_dc_left_predictor_32x8 = aom_dc_left_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x8 = aom_dc_left_predictor_32x8_sse2;
+    aom_dc_left_predictor_4x16 = aom_dc_left_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x16 = aom_dc_left_predictor_4x16_sse2;
+    aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_sse2;
+    aom_dc_left_predictor_4x8 = aom_dc_left_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x8 = aom_dc_left_predictor_4x8_sse2;
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
+    aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
+    aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_avx2;
+    aom_dc_left_predictor_8x16 = aom_dc_left_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x16 = aom_dc_left_predictor_8x16_sse2;
+    aom_dc_left_predictor_8x32 = aom_dc_left_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x32 = aom_dc_left_predictor_8x32_sse2;
+    aom_dc_left_predictor_8x4 = aom_dc_left_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x4 = aom_dc_left_predictor_8x4_sse2;
+    aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_sse2;
+    aom_dc_predictor_16x16 = aom_dc_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x16 = aom_dc_predictor_16x16_sse2;
+    aom_dc_predictor_16x32 = aom_dc_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x32 = aom_dc_predictor_16x32_sse2;
+    aom_dc_predictor_16x4 = aom_dc_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x4 = aom_dc_predictor_16x4_sse2;
+    aom_dc_predictor_16x64 = aom_dc_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x64 = aom_dc_predictor_16x64_sse2;
+    aom_dc_predictor_16x8 = aom_dc_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x8 = aom_dc_predictor_16x8_sse2;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_avx2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_avx2;
+    aom_dc_predictor_32x64 = aom_dc_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+    aom_dc_predictor_32x8 = aom_dc_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x8 = aom_dc_predictor_32x8_sse2;
+    aom_dc_predictor_4x16 = aom_dc_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x16 = aom_dc_predictor_4x16_sse2;
+    aom_dc_predictor_4x4 = aom_dc_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x4 = aom_dc_predictor_4x4_sse2;
+    aom_dc_predictor_4x8 = aom_dc_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x8 = aom_dc_predictor_4x8_sse2;
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
+    aom_dc_predictor_64x32 = aom_dc_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
+    aom_dc_predictor_64x64 = aom_dc_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_avx2;
+    aom_dc_predictor_8x16 = aom_dc_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x16 = aom_dc_predictor_8x16_sse2;
+    aom_dc_predictor_8x32 = aom_dc_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x32 = aom_dc_predictor_8x32_sse2;
+    aom_dc_predictor_8x4 = aom_dc_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x4 = aom_dc_predictor_8x4_sse2;
+    aom_dc_predictor_8x8 = aom_dc_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x8 = aom_dc_predictor_8x8_sse2;
+    aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_sse2;
+    aom_dc_top_predictor_16x32 = aom_dc_top_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x32 = aom_dc_top_predictor_16x32_sse2;
+    aom_dc_top_predictor_16x4 = aom_dc_top_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x4 = aom_dc_top_predictor_16x4_sse2;
+    aom_dc_top_predictor_16x64 = aom_dc_top_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x64 = aom_dc_top_predictor_16x64_sse2;
+    aom_dc_top_predictor_16x8 = aom_dc_top_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x8 = aom_dc_top_predictor_16x8_sse2;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_avx2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_avx2;
+    aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+    aom_dc_top_predictor_32x8 = aom_dc_top_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x8 = aom_dc_top_predictor_32x8_sse2;
+    aom_dc_top_predictor_4x16 = aom_dc_top_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x16 = aom_dc_top_predictor_4x16_sse2;
+    aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_sse2;
+    aom_dc_top_predictor_4x8 = aom_dc_top_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x8 = aom_dc_top_predictor_4x8_sse2;
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
+    aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
+    aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_avx2;
+    aom_dc_top_predictor_8x16 = aom_dc_top_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x16 = aom_dc_top_predictor_8x16_sse2;
+    aom_dc_top_predictor_8x32 = aom_dc_top_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x32 = aom_dc_top_predictor_8x32_sse2;
+    aom_dc_top_predictor_8x4 = aom_dc_top_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x4 = aom_dc_top_predictor_8x4_sse2;
+    aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_sse2;
+    aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_ssse3;
+    aom_dist_wtd_sad128x128_avg = aom_dist_wtd_sad128x128_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad128x128_avg = aom_dist_wtd_sad128x128_avg_sse2;
+    aom_dist_wtd_sad128x64_avg = aom_dist_wtd_sad128x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad128x64_avg = aom_dist_wtd_sad128x64_avg_sse2;
+    aom_dist_wtd_sad16x16_avg = aom_dist_wtd_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x16_avg = aom_dist_wtd_sad16x16_avg_sse2;
+    aom_dist_wtd_sad16x32_avg = aom_dist_wtd_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x32_avg = aom_dist_wtd_sad16x32_avg_sse2;
+    aom_dist_wtd_sad16x4_avg = aom_dist_wtd_sad16x4_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x4_avg = aom_dist_wtd_sad16x4_avg_sse2;
+    aom_dist_wtd_sad16x64_avg = aom_dist_wtd_sad16x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x64_avg = aom_dist_wtd_sad16x64_avg_sse2;
+    aom_dist_wtd_sad16x8_avg = aom_dist_wtd_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad16x8_avg = aom_dist_wtd_sad16x8_avg_sse2;
+    aom_dist_wtd_sad32x16_avg = aom_dist_wtd_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x16_avg = aom_dist_wtd_sad32x16_avg_sse2;
+    aom_dist_wtd_sad32x32_avg = aom_dist_wtd_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x32_avg = aom_dist_wtd_sad32x32_avg_sse2;
+    aom_dist_wtd_sad32x64_avg = aom_dist_wtd_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x64_avg = aom_dist_wtd_sad32x64_avg_sse2;
+    aom_dist_wtd_sad32x8_avg = aom_dist_wtd_sad32x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad32x8_avg = aom_dist_wtd_sad32x8_avg_sse2;
+    aom_dist_wtd_sad4x16_avg = aom_dist_wtd_sad4x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad4x16_avg = aom_dist_wtd_sad4x16_avg_sse2;
+    aom_dist_wtd_sad4x4_avg = aom_dist_wtd_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad4x4_avg = aom_dist_wtd_sad4x4_avg_sse2;
+    aom_dist_wtd_sad4x8_avg = aom_dist_wtd_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad4x8_avg = aom_dist_wtd_sad4x8_avg_sse2;
+    aom_dist_wtd_sad64x128_avg = aom_dist_wtd_sad64x128_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x128_avg = aom_dist_wtd_sad64x128_avg_sse2;
+    aom_dist_wtd_sad64x16_avg = aom_dist_wtd_sad64x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x16_avg = aom_dist_wtd_sad64x16_avg_sse2;
+    aom_dist_wtd_sad64x32_avg = aom_dist_wtd_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x32_avg = aom_dist_wtd_sad64x32_avg_sse2;
+    aom_dist_wtd_sad64x64_avg = aom_dist_wtd_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad64x64_avg = aom_dist_wtd_sad64x64_avg_sse2;
+    aom_dist_wtd_sad8x16_avg = aom_dist_wtd_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x16_avg = aom_dist_wtd_sad8x16_avg_sse2;
+    aom_dist_wtd_sad8x32_avg = aom_dist_wtd_sad8x32_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x32_avg = aom_dist_wtd_sad8x32_avg_sse2;
+    aom_dist_wtd_sad8x4_avg = aom_dist_wtd_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x4_avg = aom_dist_wtd_sad8x4_avg_sse2;
+    aom_dist_wtd_sad8x8_avg = aom_dist_wtd_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_dist_wtd_sad8x8_avg = aom_dist_wtd_sad8x8_avg_sse2;
+    aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3;
+    aom_fdct4x4 = aom_fdct4x4_c;
+    if (flags & HAS_SSE2) aom_fdct4x4 = aom_fdct4x4_sse2;
+    aom_fdct4x4_lp = aom_fdct4x4_lp_c;
+    if (flags & HAS_SSE2) aom_fdct4x4_lp = aom_fdct4x4_lp_sse2;
+    aom_fft16x16_float = aom_fft16x16_float_c;
+    if (flags & HAS_SSE2) aom_fft16x16_float = aom_fft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_fft16x16_float = aom_fft16x16_float_avx2;
+    aom_fft32x32_float = aom_fft32x32_float_c;
+    if (flags & HAS_SSE2) aom_fft32x32_float = aom_fft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_fft32x32_float = aom_fft32x32_float_avx2;
+    aom_fft4x4_float = aom_fft4x4_float_c;
+    if (flags & HAS_SSE2) aom_fft4x4_float = aom_fft4x4_float_sse2;
+    aom_fft8x8_float = aom_fft8x8_float_c;
+    if (flags & HAS_SSE2) aom_fft8x8_float = aom_fft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_fft8x8_float = aom_fft8x8_float_avx2;
+    aom_get_blk_sse_sum = aom_get_blk_sse_sum_c;
+    if (flags & HAS_SSE2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_sse2;
+    if (flags & HAS_AVX2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_avx2;
+    aom_get_mb_ss = aom_get_mb_ss_c;
+    if (flags & HAS_SSE2) aom_get_mb_ss = aom_get_mb_ss_sse2;
+    aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_c;
+    if (flags & HAS_SSE2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_avx2;
+    aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_c;
+    if (flags & HAS_SSE2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_avx2;
+    aom_h_predictor_16x16 = aom_h_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x16 = aom_h_predictor_16x16_sse2;
+    aom_h_predictor_16x32 = aom_h_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x32 = aom_h_predictor_16x32_sse2;
+    aom_h_predictor_16x4 = aom_h_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x4 = aom_h_predictor_16x4_sse2;
+    aom_h_predictor_16x64 = aom_h_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x64 = aom_h_predictor_16x64_sse2;
+    aom_h_predictor_16x8 = aom_h_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x8 = aom_h_predictor_16x8_sse2;
+    aom_h_predictor_32x16 = aom_h_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x16 = aom_h_predictor_32x16_sse2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_avx2;
+    aom_h_predictor_32x64 = aom_h_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x64 = aom_h_predictor_32x64_sse2;
+    aom_h_predictor_32x8 = aom_h_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x8 = aom_h_predictor_32x8_sse2;
+    aom_h_predictor_4x16 = aom_h_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x16 = aom_h_predictor_4x16_sse2;
+    aom_h_predictor_4x4 = aom_h_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x4 = aom_h_predictor_4x4_sse2;
+    aom_h_predictor_4x8 = aom_h_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x8 = aom_h_predictor_4x8_sse2;
+    aom_h_predictor_64x16 = aom_h_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_64x16 = aom_h_predictor_64x16_sse2;
+    aom_h_predictor_64x32 = aom_h_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_64x32 = aom_h_predictor_64x32_sse2;
+    aom_h_predictor_64x64 = aom_h_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_h_predictor_64x64 = aom_h_predictor_64x64_sse2;
+    aom_h_predictor_8x16 = aom_h_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x16 = aom_h_predictor_8x16_sse2;
+    aom_h_predictor_8x32 = aom_h_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x32 = aom_h_predictor_8x32_sse2;
+    aom_h_predictor_8x4 = aom_h_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x4 = aom_h_predictor_8x4_sse2;
+    aom_h_predictor_8x8 = aom_h_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x8 = aom_h_predictor_8x8_sse2;
+    aom_hadamard_16x16 = aom_hadamard_16x16_c;
+    if (flags & HAS_SSE2) aom_hadamard_16x16 = aom_hadamard_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_16x16 = aom_hadamard_16x16_avx2;
+    aom_hadamard_32x32 = aom_hadamard_32x32_c;
+    if (flags & HAS_SSE2) aom_hadamard_32x32 = aom_hadamard_32x32_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_32x32 = aom_hadamard_32x32_avx2;
+    aom_hadamard_4x4 = aom_hadamard_4x4_c;
+    if (flags & HAS_SSE2) aom_hadamard_4x4 = aom_hadamard_4x4_sse2;
+    aom_hadamard_8x8 = aom_hadamard_8x8_c;
+    if (flags & HAS_SSE2) aom_hadamard_8x8 = aom_hadamard_8x8_sse2;
+    aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_c;
+    if (flags & HAS_SSE2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_avx2;
+    aom_hadamard_lp_8x8 = aom_hadamard_lp_8x8_c;
+    if (flags & HAS_SSE2) aom_hadamard_lp_8x8 = aom_hadamard_lp_8x8_sse2;
+    aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_c;
+    if (flags & HAS_SSE2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_avx2;
+    aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_sse2;
+    aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
+    aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_sse4_1;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_sse4_1;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_sse4_1;
+    aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_sse4_1;
+    aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_sse4_1;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_sse4_1;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_sse4_1;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_sse4_1;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_sse4_1;
+    aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_sse4_1;
+    aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_sse4_1;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_sse4_1;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_sse4_1;
+    aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_sse4_1;
+    aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_sse4_1;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_sse4_1;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_sse4_1;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_sse4_1;
+    aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_sse4_1;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_sse4_1;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance16x16 = aom_highbd_10_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x16 = aom_highbd_10_sub_pixel_avg_variance16x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x32 = aom_highbd_10_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x32 = aom_highbd_10_sub_pixel_avg_variance16x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x4 = aom_highbd_10_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x4 = aom_highbd_10_sub_pixel_avg_variance16x4_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x64 = aom_highbd_10_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x64 = aom_highbd_10_sub_pixel_avg_variance16x64_sse2;
+    aom_highbd_10_sub_pixel_avg_variance16x8 = aom_highbd_10_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance16x8 = aom_highbd_10_sub_pixel_avg_variance16x8_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x16 = aom_highbd_10_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x16 = aom_highbd_10_sub_pixel_avg_variance32x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x32 = aom_highbd_10_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x32 = aom_highbd_10_sub_pixel_avg_variance32x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x64 = aom_highbd_10_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x64 = aom_highbd_10_sub_pixel_avg_variance32x64_sse2;
+    aom_highbd_10_sub_pixel_avg_variance32x8 = aom_highbd_10_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance32x8 = aom_highbd_10_sub_pixel_avg_variance32x8_sse2;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance64x16 = aom_highbd_10_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance64x16 = aom_highbd_10_sub_pixel_avg_variance64x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance64x32 = aom_highbd_10_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance64x32 = aom_highbd_10_sub_pixel_avg_variance64x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance64x64 = aom_highbd_10_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance64x64 = aom_highbd_10_sub_pixel_avg_variance64x64_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x16 = aom_highbd_10_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x16 = aom_highbd_10_sub_pixel_avg_variance8x16_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x32 = aom_highbd_10_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x32 = aom_highbd_10_sub_pixel_avg_variance8x32_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x4 = aom_highbd_10_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x4 = aom_highbd_10_sub_pixel_avg_variance8x4_sse2;
+    aom_highbd_10_sub_pixel_avg_variance8x8 = aom_highbd_10_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_avg_variance8x8 = aom_highbd_10_sub_pixel_avg_variance8x8_sse2;
+    aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_avx2;
+    aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_avx2;
+    aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_avx2;
+    aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_avx2;
+    aom_highbd_10_sub_pixel_variance16x4 = aom_highbd_10_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x4 = aom_highbd_10_sub_pixel_variance16x4_sse2;
+    aom_highbd_10_sub_pixel_variance16x64 = aom_highbd_10_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x64 = aom_highbd_10_sub_pixel_variance16x64_sse2;
+    aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_avx2;
+    aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_avx2;
+    aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_avx2;
+    aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_avx2;
+    aom_highbd_10_sub_pixel_variance32x8 = aom_highbd_10_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance32x8 = aom_highbd_10_sub_pixel_variance32x8_sse2;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_avx2;
+    aom_highbd_10_sub_pixel_variance64x16 = aom_highbd_10_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x16 = aom_highbd_10_sub_pixel_variance64x16_sse2;
+    aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_avx2;
+    aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_avx2;
+    aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_avx2;
+    aom_highbd_10_sub_pixel_variance8x32 = aom_highbd_10_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x32 = aom_highbd_10_sub_pixel_variance8x32_sse2;
+    aom_highbd_10_sub_pixel_variance8x4 = aom_highbd_10_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x4 = aom_highbd_10_sub_pixel_variance8x4_sse2;
+    aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_avx2;
+    aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_avx2;
+    aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_avx2;
+    aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_avx2;
+    aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_avx2;
+    aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_avx2;
+    aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_avx2;
+    aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_avx2;
+    aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_avx2;
+    aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_avx2;
+    aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_avx2;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_sse4_1;
+    aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_avx2;
+    aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_avx2;
+    aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_avx2;
+    aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_avx2;
+    aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_avx2;
+    aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_avx2;
+    aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_avx2;
+    aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_12_mse16x16 = aom_highbd_12_mse16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_mse16x16 = aom_highbd_12_mse16x16_sse2;
+    aom_highbd_12_mse8x8 = aom_highbd_12_mse8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_mse8x8 = aom_highbd_12_mse8x8_sse2;
+    aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_sse4_1;
+    aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_sse4_1;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_sse4_1;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_sse4_1;
+    aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_sse4_1;
+    aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_sse4_1;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_sse4_1;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_sse4_1;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_sse4_1;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_sse4_1;
+    aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_sse4_1;
+    aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_sse4_1;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_sse4_1;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_sse4_1;
+    aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_sse4_1;
+    aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_sse4_1;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_sse4_1;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_sse4_1;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_sse4_1;
+    aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_sse4_1;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_sse4_1;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance16x16 = aom_highbd_12_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x16 = aom_highbd_12_sub_pixel_avg_variance16x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x32 = aom_highbd_12_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x32 = aom_highbd_12_sub_pixel_avg_variance16x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x4 = aom_highbd_12_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x4 = aom_highbd_12_sub_pixel_avg_variance16x4_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x64 = aom_highbd_12_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x64 = aom_highbd_12_sub_pixel_avg_variance16x64_sse2;
+    aom_highbd_12_sub_pixel_avg_variance16x8 = aom_highbd_12_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance16x8 = aom_highbd_12_sub_pixel_avg_variance16x8_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x16 = aom_highbd_12_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x16 = aom_highbd_12_sub_pixel_avg_variance32x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x32 = aom_highbd_12_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x32 = aom_highbd_12_sub_pixel_avg_variance32x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x64 = aom_highbd_12_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x64 = aom_highbd_12_sub_pixel_avg_variance32x64_sse2;
+    aom_highbd_12_sub_pixel_avg_variance32x8 = aom_highbd_12_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance32x8 = aom_highbd_12_sub_pixel_avg_variance32x8_sse2;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance64x16 = aom_highbd_12_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance64x16 = aom_highbd_12_sub_pixel_avg_variance64x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance64x32 = aom_highbd_12_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance64x32 = aom_highbd_12_sub_pixel_avg_variance64x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance64x64 = aom_highbd_12_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance64x64 = aom_highbd_12_sub_pixel_avg_variance64x64_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x16 = aom_highbd_12_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x16 = aom_highbd_12_sub_pixel_avg_variance8x16_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x32 = aom_highbd_12_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x32 = aom_highbd_12_sub_pixel_avg_variance8x32_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x4 = aom_highbd_12_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x4 = aom_highbd_12_sub_pixel_avg_variance8x4_sse2;
+    aom_highbd_12_sub_pixel_avg_variance8x8 = aom_highbd_12_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_avg_variance8x8 = aom_highbd_12_sub_pixel_avg_variance8x8_sse2;
+    aom_highbd_12_sub_pixel_variance128x128 = aom_highbd_12_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance128x128 = aom_highbd_12_sub_pixel_variance128x128_sse2;
+    aom_highbd_12_sub_pixel_variance128x64 = aom_highbd_12_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance128x64 = aom_highbd_12_sub_pixel_variance128x64_sse2;
+    aom_highbd_12_sub_pixel_variance16x16 = aom_highbd_12_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x16 = aom_highbd_12_sub_pixel_variance16x16_sse2;
+    aom_highbd_12_sub_pixel_variance16x32 = aom_highbd_12_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x32 = aom_highbd_12_sub_pixel_variance16x32_sse2;
+    aom_highbd_12_sub_pixel_variance16x4 = aom_highbd_12_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x4 = aom_highbd_12_sub_pixel_variance16x4_sse2;
+    aom_highbd_12_sub_pixel_variance16x64 = aom_highbd_12_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x64 = aom_highbd_12_sub_pixel_variance16x64_sse2;
+    aom_highbd_12_sub_pixel_variance16x8 = aom_highbd_12_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance16x8 = aom_highbd_12_sub_pixel_variance16x8_sse2;
+    aom_highbd_12_sub_pixel_variance32x16 = aom_highbd_12_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x16 = aom_highbd_12_sub_pixel_variance32x16_sse2;
+    aom_highbd_12_sub_pixel_variance32x32 = aom_highbd_12_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x32 = aom_highbd_12_sub_pixel_variance32x32_sse2;
+    aom_highbd_12_sub_pixel_variance32x64 = aom_highbd_12_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x64 = aom_highbd_12_sub_pixel_variance32x64_sse2;
+    aom_highbd_12_sub_pixel_variance32x8 = aom_highbd_12_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance32x8 = aom_highbd_12_sub_pixel_variance32x8_sse2;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_variance64x128 = aom_highbd_12_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x128 = aom_highbd_12_sub_pixel_variance64x128_sse2;
+    aom_highbd_12_sub_pixel_variance64x16 = aom_highbd_12_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x16 = aom_highbd_12_sub_pixel_variance64x16_sse2;
+    aom_highbd_12_sub_pixel_variance64x32 = aom_highbd_12_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x32 = aom_highbd_12_sub_pixel_variance64x32_sse2;
+    aom_highbd_12_sub_pixel_variance64x64 = aom_highbd_12_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance64x64 = aom_highbd_12_sub_pixel_variance64x64_sse2;
+    aom_highbd_12_sub_pixel_variance8x16 = aom_highbd_12_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x16 = aom_highbd_12_sub_pixel_variance8x16_sse2;
+    aom_highbd_12_sub_pixel_variance8x32 = aom_highbd_12_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x32 = aom_highbd_12_sub_pixel_variance8x32_sse2;
+    aom_highbd_12_sub_pixel_variance8x4 = aom_highbd_12_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x4 = aom_highbd_12_sub_pixel_variance8x4_sse2;
+    aom_highbd_12_sub_pixel_variance8x8 = aom_highbd_12_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_sub_pixel_variance8x8 = aom_highbd_12_sub_pixel_variance8x8_sse2;
+    aom_highbd_12_variance128x128 = aom_highbd_12_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance128x128 = aom_highbd_12_variance128x128_sse2;
+    aom_highbd_12_variance128x64 = aom_highbd_12_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance128x64 = aom_highbd_12_variance128x64_sse2;
+    aom_highbd_12_variance16x16 = aom_highbd_12_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x16 = aom_highbd_12_variance16x16_sse2;
+    aom_highbd_12_variance16x32 = aom_highbd_12_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x32 = aom_highbd_12_variance16x32_sse2;
+    aom_highbd_12_variance16x64 = aom_highbd_12_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x64 = aom_highbd_12_variance16x64_sse2;
+    aom_highbd_12_variance16x8 = aom_highbd_12_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance16x8 = aom_highbd_12_variance16x8_sse2;
+    aom_highbd_12_variance32x16 = aom_highbd_12_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x16 = aom_highbd_12_variance32x16_sse2;
+    aom_highbd_12_variance32x32 = aom_highbd_12_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x32 = aom_highbd_12_variance32x32_sse2;
+    aom_highbd_12_variance32x64 = aom_highbd_12_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x64 = aom_highbd_12_variance32x64_sse2;
+    aom_highbd_12_variance32x8 = aom_highbd_12_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance32x8 = aom_highbd_12_variance32x8_sse2;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_sse4_1;
+    aom_highbd_12_variance64x128 = aom_highbd_12_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x128 = aom_highbd_12_variance64x128_sse2;
+    aom_highbd_12_variance64x16 = aom_highbd_12_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x16 = aom_highbd_12_variance64x16_sse2;
+    aom_highbd_12_variance64x32 = aom_highbd_12_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x32 = aom_highbd_12_variance64x32_sse2;
+    aom_highbd_12_variance64x64 = aom_highbd_12_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance64x64 = aom_highbd_12_variance64x64_sse2;
+    aom_highbd_12_variance8x16 = aom_highbd_12_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance8x16 = aom_highbd_12_variance8x16_sse2;
+    aom_highbd_12_variance8x32 = aom_highbd_12_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance8x32 = aom_highbd_12_variance8x32_sse2;
+    aom_highbd_12_variance8x8 = aom_highbd_12_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_12_variance8x8 = aom_highbd_12_variance8x8_sse2;
+    aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_8_mse16x16 = aom_highbd_8_mse16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_mse16x16 = aom_highbd_8_mse16x16_sse2;
+    aom_highbd_8_mse8x8 = aom_highbd_8_mse8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_mse8x8 = aom_highbd_8_mse8x8_sse2;
+    aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_sse4_1;
+    aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_sse4_1;
+    aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_sse4_1;
+    aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_sse4_1;
+    aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_sse4_1;
+    aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_sse4_1;
+    aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_sse4_1;
+    aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_sse4_1;
+    aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_sse4_1;
+    aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_sse4_1;
+    aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_sse4_1;
+    aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_sse4_1;
+    aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_sse4_1;
+    aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_sse4_1;
+    aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_sse4_1;
+    aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_sse4_1;
+    aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_sse4_1;
+    aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_sse4_1;
+    aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_sse4_1;
+    aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_sse4_1;
+    aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_sse4_1;
+    aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance16x16 = aom_highbd_8_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x16 = aom_highbd_8_sub_pixel_avg_variance16x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x32 = aom_highbd_8_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x32 = aom_highbd_8_sub_pixel_avg_variance16x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x4 = aom_highbd_8_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x4 = aom_highbd_8_sub_pixel_avg_variance16x4_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x64 = aom_highbd_8_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x64 = aom_highbd_8_sub_pixel_avg_variance16x64_sse2;
+    aom_highbd_8_sub_pixel_avg_variance16x8 = aom_highbd_8_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance16x8 = aom_highbd_8_sub_pixel_avg_variance16x8_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x16 = aom_highbd_8_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x16 = aom_highbd_8_sub_pixel_avg_variance32x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x32 = aom_highbd_8_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x32 = aom_highbd_8_sub_pixel_avg_variance32x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x64 = aom_highbd_8_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x64 = aom_highbd_8_sub_pixel_avg_variance32x64_sse2;
+    aom_highbd_8_sub_pixel_avg_variance32x8 = aom_highbd_8_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance32x8 = aom_highbd_8_sub_pixel_avg_variance32x8_sse2;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance64x16 = aom_highbd_8_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance64x16 = aom_highbd_8_sub_pixel_avg_variance64x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance64x32 = aom_highbd_8_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance64x32 = aom_highbd_8_sub_pixel_avg_variance64x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance64x64 = aom_highbd_8_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance64x64 = aom_highbd_8_sub_pixel_avg_variance64x64_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x16 = aom_highbd_8_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x16 = aom_highbd_8_sub_pixel_avg_variance8x16_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x32 = aom_highbd_8_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x32 = aom_highbd_8_sub_pixel_avg_variance8x32_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x4 = aom_highbd_8_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x4 = aom_highbd_8_sub_pixel_avg_variance8x4_sse2;
+    aom_highbd_8_sub_pixel_avg_variance8x8 = aom_highbd_8_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_avg_variance8x8 = aom_highbd_8_sub_pixel_avg_variance8x8_sse2;
+    aom_highbd_8_sub_pixel_variance128x128 = aom_highbd_8_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance128x128 = aom_highbd_8_sub_pixel_variance128x128_sse2;
+    aom_highbd_8_sub_pixel_variance128x64 = aom_highbd_8_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance128x64 = aom_highbd_8_sub_pixel_variance128x64_sse2;
+    aom_highbd_8_sub_pixel_variance16x16 = aom_highbd_8_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x16 = aom_highbd_8_sub_pixel_variance16x16_sse2;
+    aom_highbd_8_sub_pixel_variance16x32 = aom_highbd_8_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x32 = aom_highbd_8_sub_pixel_variance16x32_sse2;
+    aom_highbd_8_sub_pixel_variance16x4 = aom_highbd_8_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x4 = aom_highbd_8_sub_pixel_variance16x4_sse2;
+    aom_highbd_8_sub_pixel_variance16x64 = aom_highbd_8_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x64 = aom_highbd_8_sub_pixel_variance16x64_sse2;
+    aom_highbd_8_sub_pixel_variance16x8 = aom_highbd_8_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance16x8 = aom_highbd_8_sub_pixel_variance16x8_sse2;
+    aom_highbd_8_sub_pixel_variance32x16 = aom_highbd_8_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x16 = aom_highbd_8_sub_pixel_variance32x16_sse2;
+    aom_highbd_8_sub_pixel_variance32x32 = aom_highbd_8_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x32 = aom_highbd_8_sub_pixel_variance32x32_sse2;
+    aom_highbd_8_sub_pixel_variance32x64 = aom_highbd_8_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x64 = aom_highbd_8_sub_pixel_variance32x64_sse2;
+    aom_highbd_8_sub_pixel_variance32x8 = aom_highbd_8_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance32x8 = aom_highbd_8_sub_pixel_variance32x8_sse2;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_variance64x128 = aom_highbd_8_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x128 = aom_highbd_8_sub_pixel_variance64x128_sse2;
+    aom_highbd_8_sub_pixel_variance64x16 = aom_highbd_8_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x16 = aom_highbd_8_sub_pixel_variance64x16_sse2;
+    aom_highbd_8_sub_pixel_variance64x32 = aom_highbd_8_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x32 = aom_highbd_8_sub_pixel_variance64x32_sse2;
+    aom_highbd_8_sub_pixel_variance64x64 = aom_highbd_8_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance64x64 = aom_highbd_8_sub_pixel_variance64x64_sse2;
+    aom_highbd_8_sub_pixel_variance8x16 = aom_highbd_8_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x16 = aom_highbd_8_sub_pixel_variance8x16_sse2;
+    aom_highbd_8_sub_pixel_variance8x32 = aom_highbd_8_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x32 = aom_highbd_8_sub_pixel_variance8x32_sse2;
+    aom_highbd_8_sub_pixel_variance8x4 = aom_highbd_8_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x4 = aom_highbd_8_sub_pixel_variance8x4_sse2;
+    aom_highbd_8_sub_pixel_variance8x8 = aom_highbd_8_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_sub_pixel_variance8x8 = aom_highbd_8_sub_pixel_variance8x8_sse2;
+    aom_highbd_8_variance128x128 = aom_highbd_8_variance128x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance128x128 = aom_highbd_8_variance128x128_sse2;
+    aom_highbd_8_variance128x64 = aom_highbd_8_variance128x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance128x64 = aom_highbd_8_variance128x64_sse2;
+    aom_highbd_8_variance16x16 = aom_highbd_8_variance16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x16 = aom_highbd_8_variance16x16_sse2;
+    aom_highbd_8_variance16x32 = aom_highbd_8_variance16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x32 = aom_highbd_8_variance16x32_sse2;
+    aom_highbd_8_variance16x64 = aom_highbd_8_variance16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x64 = aom_highbd_8_variance16x64_sse2;
+    aom_highbd_8_variance16x8 = aom_highbd_8_variance16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance16x8 = aom_highbd_8_variance16x8_sse2;
+    aom_highbd_8_variance32x16 = aom_highbd_8_variance32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x16 = aom_highbd_8_variance32x16_sse2;
+    aom_highbd_8_variance32x32 = aom_highbd_8_variance32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x32 = aom_highbd_8_variance32x32_sse2;
+    aom_highbd_8_variance32x64 = aom_highbd_8_variance32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x64 = aom_highbd_8_variance32x64_sse2;
+    aom_highbd_8_variance32x8 = aom_highbd_8_variance32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance32x8 = aom_highbd_8_variance32x8_sse2;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_sse4_1;
+    aom_highbd_8_variance64x128 = aom_highbd_8_variance64x128_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x128 = aom_highbd_8_variance64x128_sse2;
+    aom_highbd_8_variance64x16 = aom_highbd_8_variance64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x16 = aom_highbd_8_variance64x16_sse2;
+    aom_highbd_8_variance64x32 = aom_highbd_8_variance64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x32 = aom_highbd_8_variance64x32_sse2;
+    aom_highbd_8_variance64x64 = aom_highbd_8_variance64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance64x64 = aom_highbd_8_variance64x64_sse2;
+    aom_highbd_8_variance8x16 = aom_highbd_8_variance8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance8x16 = aom_highbd_8_variance8x16_sse2;
+    aom_highbd_8_variance8x32 = aom_highbd_8_variance8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance8x32 = aom_highbd_8_variance8x32_sse2;
+    aom_highbd_8_variance8x8 = aom_highbd_8_variance8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_8_variance8x8 = aom_highbd_8_variance8x8_sse2;
+    aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_avx2;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_sse4_1;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_sse4_1;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
+    aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_sse2;
+    if (flags & HAS_AVX2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_avx2;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_c;
+    if (flags & HAS_SSE2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_avx2;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_c;
+    if (flags & HAS_SSE2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_avx2;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_c;
+    if (flags & HAS_SSE2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_dc_128_predictor_16x16 = aom_highbd_dc_128_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_16x16 = aom_highbd_dc_128_predictor_16x16_sse2;
+    aom_highbd_dc_128_predictor_16x32 = aom_highbd_dc_128_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_16x32 = aom_highbd_dc_128_predictor_16x32_sse2;
+    aom_highbd_dc_128_predictor_16x8 = aom_highbd_dc_128_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_16x8 = aom_highbd_dc_128_predictor_16x8_sse2;
+    aom_highbd_dc_128_predictor_32x16 = aom_highbd_dc_128_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_32x16 = aom_highbd_dc_128_predictor_32x16_sse2;
+    aom_highbd_dc_128_predictor_32x32 = aom_highbd_dc_128_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_32x32 = aom_highbd_dc_128_predictor_32x32_sse2;
+    aom_highbd_dc_128_predictor_4x4 = aom_highbd_dc_128_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_4x4 = aom_highbd_dc_128_predictor_4x4_sse2;
+    aom_highbd_dc_128_predictor_4x8 = aom_highbd_dc_128_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_4x8 = aom_highbd_dc_128_predictor_4x8_sse2;
+    aom_highbd_dc_128_predictor_8x16 = aom_highbd_dc_128_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_8x16 = aom_highbd_dc_128_predictor_8x16_sse2;
+    aom_highbd_dc_128_predictor_8x4 = aom_highbd_dc_128_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_8x4 = aom_highbd_dc_128_predictor_8x4_sse2;
+    aom_highbd_dc_128_predictor_8x8 = aom_highbd_dc_128_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_128_predictor_8x8 = aom_highbd_dc_128_predictor_8x8_sse2;
+    aom_highbd_dc_left_predictor_16x16 = aom_highbd_dc_left_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_16x16 = aom_highbd_dc_left_predictor_16x16_sse2;
+    aom_highbd_dc_left_predictor_16x32 = aom_highbd_dc_left_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_16x32 = aom_highbd_dc_left_predictor_16x32_sse2;
+    aom_highbd_dc_left_predictor_16x8 = aom_highbd_dc_left_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_16x8 = aom_highbd_dc_left_predictor_16x8_sse2;
+    aom_highbd_dc_left_predictor_32x16 = aom_highbd_dc_left_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_32x16 = aom_highbd_dc_left_predictor_32x16_sse2;
+    aom_highbd_dc_left_predictor_32x32 = aom_highbd_dc_left_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_32x32 = aom_highbd_dc_left_predictor_32x32_sse2;
+    aom_highbd_dc_left_predictor_4x4 = aom_highbd_dc_left_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_4x4 = aom_highbd_dc_left_predictor_4x4_sse2;
+    aom_highbd_dc_left_predictor_4x8 = aom_highbd_dc_left_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_4x8 = aom_highbd_dc_left_predictor_4x8_sse2;
+    aom_highbd_dc_left_predictor_8x16 = aom_highbd_dc_left_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_8x16 = aom_highbd_dc_left_predictor_8x16_sse2;
+    aom_highbd_dc_left_predictor_8x4 = aom_highbd_dc_left_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_8x4 = aom_highbd_dc_left_predictor_8x4_sse2;
+    aom_highbd_dc_left_predictor_8x8 = aom_highbd_dc_left_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_left_predictor_8x8 = aom_highbd_dc_left_predictor_8x8_sse2;
+    aom_highbd_dc_predictor_16x16 = aom_highbd_dc_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_16x16 = aom_highbd_dc_predictor_16x16_sse2;
+    aom_highbd_dc_predictor_16x32 = aom_highbd_dc_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_16x32 = aom_highbd_dc_predictor_16x32_sse2;
+    aom_highbd_dc_predictor_16x8 = aom_highbd_dc_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_16x8 = aom_highbd_dc_predictor_16x8_sse2;
+    aom_highbd_dc_predictor_32x16 = aom_highbd_dc_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_32x16 = aom_highbd_dc_predictor_32x16_sse2;
+    aom_highbd_dc_predictor_32x32 = aom_highbd_dc_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_32x32 = aom_highbd_dc_predictor_32x32_sse2;
+    aom_highbd_dc_predictor_4x4 = aom_highbd_dc_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_4x4 = aom_highbd_dc_predictor_4x4_sse2;
+    aom_highbd_dc_predictor_4x8 = aom_highbd_dc_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_4x8 = aom_highbd_dc_predictor_4x8_sse2;
+    aom_highbd_dc_predictor_8x16 = aom_highbd_dc_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_8x16 = aom_highbd_dc_predictor_8x16_sse2;
+    aom_highbd_dc_predictor_8x4 = aom_highbd_dc_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_8x4 = aom_highbd_dc_predictor_8x4_sse2;
+    aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_sse2;
+    aom_highbd_dc_top_predictor_16x16 = aom_highbd_dc_top_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_16x16 = aom_highbd_dc_top_predictor_16x16_sse2;
+    aom_highbd_dc_top_predictor_16x32 = aom_highbd_dc_top_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_16x32 = aom_highbd_dc_top_predictor_16x32_sse2;
+    aom_highbd_dc_top_predictor_16x8 = aom_highbd_dc_top_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_16x8 = aom_highbd_dc_top_predictor_16x8_sse2;
+    aom_highbd_dc_top_predictor_32x16 = aom_highbd_dc_top_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_32x16 = aom_highbd_dc_top_predictor_32x16_sse2;
+    aom_highbd_dc_top_predictor_32x32 = aom_highbd_dc_top_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_32x32 = aom_highbd_dc_top_predictor_32x32_sse2;
+    aom_highbd_dc_top_predictor_4x4 = aom_highbd_dc_top_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_4x4 = aom_highbd_dc_top_predictor_4x4_sse2;
+    aom_highbd_dc_top_predictor_4x8 = aom_highbd_dc_top_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_4x8 = aom_highbd_dc_top_predictor_4x8_sse2;
+    aom_highbd_dc_top_predictor_8x16 = aom_highbd_dc_top_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_8x16 = aom_highbd_dc_top_predictor_8x16_sse2;
+    aom_highbd_dc_top_predictor_8x4 = aom_highbd_dc_top_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_8x4 = aom_highbd_dc_top_predictor_8x4_sse2;
+    aom_highbd_dc_top_predictor_8x8 = aom_highbd_dc_top_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_dc_top_predictor_8x8 = aom_highbd_dc_top_predictor_8x8_sse2;
+    aom_highbd_dist_wtd_comp_avg_pred = aom_highbd_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_dist_wtd_comp_avg_pred = aom_highbd_dist_wtd_comp_avg_pred_sse2;
+    aom_highbd_h_predictor_16x16 = aom_highbd_h_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_16x16 = aom_highbd_h_predictor_16x16_sse2;
+    aom_highbd_h_predictor_16x32 = aom_highbd_h_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_16x32 = aom_highbd_h_predictor_16x32_sse2;
+    aom_highbd_h_predictor_16x8 = aom_highbd_h_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_16x8 = aom_highbd_h_predictor_16x8_sse2;
+    aom_highbd_h_predictor_32x16 = aom_highbd_h_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_32x16 = aom_highbd_h_predictor_32x16_sse2;
+    aom_highbd_h_predictor_32x32 = aom_highbd_h_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_32x32 = aom_highbd_h_predictor_32x32_sse2;
+    aom_highbd_h_predictor_4x4 = aom_highbd_h_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_4x4 = aom_highbd_h_predictor_4x4_sse2;
+    aom_highbd_h_predictor_4x8 = aom_highbd_h_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_4x8 = aom_highbd_h_predictor_4x8_sse2;
+    aom_highbd_h_predictor_8x16 = aom_highbd_h_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_8x16 = aom_highbd_h_predictor_8x16_sse2;
+    aom_highbd_h_predictor_8x4 = aom_highbd_h_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_8x4 = aom_highbd_h_predictor_8x4_sse2;
+    aom_highbd_h_predictor_8x8 = aom_highbd_h_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_h_predictor_8x8 = aom_highbd_h_predictor_8x8_sse2;
+    aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_avx2;
+    aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_avx2;
+    aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_avx2;
+    aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_sse2;
+    aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_avx2;
+    aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_sse2;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_avx2;
+    aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_sse2;
+    aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_sse2;
+    aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_sse2;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_avx2;
+    aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_sse2;
+    aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_avx2;
+    aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_sse2;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_avx2;
+    aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_sse2;
+    aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_sse2;
+    aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_sse2;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_c;
+    if (flags & HAS_SSE2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
+    aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_avx2;
+    aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_avx2;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_avx2;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_avx2;
+    aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_avx2;
+    aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_avx2;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_avx2;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_avx2;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_avx2;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_avx2;
+    aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_avx2;
+    aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_avx2;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_avx2;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_avx2;
+    aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_avx2;
+    aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_avx2;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_avx2;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_avx2;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_avx2;
+    aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_avx2;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_avx2;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_avx2;
+    aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_avx2;
+    aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_avx2;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_avx2;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_avx2;
+    aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_avx2;
+    aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_avx2;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_avx2;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_avx2;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_avx2;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_avx2;
+    aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_avx2;
+    aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_avx2;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_avx2;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_avx2;
+    aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_avx2;
+    aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_avx2;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_avx2;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_avx2;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_avx2;
+    aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_avx2;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_avx2;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_avx2;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b = aom_highbd_quantize_b_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b = aom_highbd_quantize_b_avx2;
+    aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_avx2;
+    aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_avx2;
+    aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_avx2;
+    aom_highbd_quantize_b_64x64_adaptive = aom_highbd_quantize_b_64x64_adaptive_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_64x64_adaptive = aom_highbd_quantize_b_64x64_adaptive_sse2;
+    aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_c;
+    if (flags & HAS_SSE2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_avx2;
+    aom_highbd_sad128x128 = aom_highbd_sad128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128 = aom_highbd_sad128x128_avx2;
+    aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_avx2;
+    aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_avx2;
+    aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_avx2;
+    aom_highbd_sad128x64 = aom_highbd_sad128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64 = aom_highbd_sad128x64_avx2;
+    aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_avx2;
+    aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_avx2;
+    aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_avx2;
+    aom_highbd_sad16x16 = aom_highbd_sad16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x16 = aom_highbd_sad16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16 = aom_highbd_sad16x16_avx2;
+    aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_avx2;
+    aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_avx2;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_avx2;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x32 = aom_highbd_sad16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32 = aom_highbd_sad16x32_avx2;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_avx2;
+    aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_avx2;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_avx2;
+    aom_highbd_sad16x4 = aom_highbd_sad16x4_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x4 = aom_highbd_sad16x4_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4 = aom_highbd_sad16x4_avx2;
+    aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_avx2;
+    aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_avx2;
+    aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_avx2;
+    aom_highbd_sad16x64 = aom_highbd_sad16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x64 = aom_highbd_sad16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64 = aom_highbd_sad16x64_avx2;
+    aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_avx2;
+    aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_avx2;
+    aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_avx2;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x8 = aom_highbd_sad16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8 = aom_highbd_sad16x8_avx2;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_avx2;
+    aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_avx2;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_avx2;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x16 = aom_highbd_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16 = aom_highbd_sad32x16_avx2;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_avx2;
+    aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_avx2;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_avx2;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x32 = aom_highbd_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32 = aom_highbd_sad32x32_avx2;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_avx2;
+    aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_avx2;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_avx2;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x64 = aom_highbd_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64 = aom_highbd_sad32x64_avx2;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_avx2;
+    aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_avx2;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_avx2;
+    aom_highbd_sad32x8 = aom_highbd_sad32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x8 = aom_highbd_sad32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8 = aom_highbd_sad32x8_avx2;
+    aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_avx2;
+    aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_avx2;
+    aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_avx2;
+    aom_highbd_sad4x16 = aom_highbd_sad4x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x16 = aom_highbd_sad4x16_sse2;
+    aom_highbd_sad4x16_avg = aom_highbd_sad4x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x16_avg = aom_highbd_sad4x16_avg_sse2;
+    aom_highbd_sad4x16x4d = aom_highbd_sad4x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x16x4d = aom_highbd_sad4x16x4d_sse2;
+    aom_highbd_sad4x4 = aom_highbd_sad4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x4 = aom_highbd_sad4x4_sse2;
+    aom_highbd_sad4x4_avg = aom_highbd_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x4_avg = aom_highbd_sad4x4_avg_sse2;
+    aom_highbd_sad4x4x4d = aom_highbd_sad4x4x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x4x4d = aom_highbd_sad4x4x4d_sse2;
+    aom_highbd_sad4x8 = aom_highbd_sad4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x8 = aom_highbd_sad4x8_sse2;
+    aom_highbd_sad4x8_avg = aom_highbd_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x8_avg = aom_highbd_sad4x8_avg_sse2;
+    aom_highbd_sad4x8x4d = aom_highbd_sad4x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad4x8x4d = aom_highbd_sad4x8x4d_sse2;
+    aom_highbd_sad64x128 = aom_highbd_sad64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128 = aom_highbd_sad64x128_avx2;
+    aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_avx2;
+    aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_avx2;
+    aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_avx2;
+    aom_highbd_sad64x16 = aom_highbd_sad64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x16 = aom_highbd_sad64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16 = aom_highbd_sad64x16_avx2;
+    aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_avx2;
+    aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_avx2;
+    aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_avx2;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x32 = aom_highbd_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32 = aom_highbd_sad64x32_avx2;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_avx2;
+    aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_avx2;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_avx2;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x64 = aom_highbd_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64 = aom_highbd_sad64x64_avx2;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_avx2;
+    aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_avx2;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_avx2;
+    aom_highbd_sad8x16 = aom_highbd_sad8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x16 = aom_highbd_sad8x16_sse2;
+    aom_highbd_sad8x16_avg = aom_highbd_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x16_avg = aom_highbd_sad8x16_avg_sse2;
+    aom_highbd_sad8x16x4d = aom_highbd_sad8x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x16x4d = aom_highbd_sad8x16x4d_sse2;
+    aom_highbd_sad8x32 = aom_highbd_sad8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x32 = aom_highbd_sad8x32_sse2;
+    aom_highbd_sad8x32_avg = aom_highbd_sad8x32_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x32_avg = aom_highbd_sad8x32_avg_sse2;
+    aom_highbd_sad8x32x4d = aom_highbd_sad8x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x32x4d = aom_highbd_sad8x32x4d_sse2;
+    aom_highbd_sad8x4 = aom_highbd_sad8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x4 = aom_highbd_sad8x4_sse2;
+    aom_highbd_sad8x4_avg = aom_highbd_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x4_avg = aom_highbd_sad8x4_avg_sse2;
+    aom_highbd_sad8x4x4d = aom_highbd_sad8x4x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x4x4d = aom_highbd_sad8x4x4d_sse2;
+    aom_highbd_sad8x8 = aom_highbd_sad8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x8 = aom_highbd_sad8x8_sse2;
+    aom_highbd_sad8x8_avg = aom_highbd_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x8_avg = aom_highbd_sad8x8_avg_sse2;
+    aom_highbd_sad8x8x4d = aom_highbd_sad8x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad8x8x4d = aom_highbd_sad8x8x4d_sse2;
+    aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_avx2;
+    aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_avx2;
+    aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_avx2;
+    aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_avx2;
+    aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_avx2;
+    aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_avx2;
+    aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_avx2;
+    aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_avx2;
+    aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_avx2;
+    aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_avx2;
+    aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_avx2;
+    aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_avx2;
+    aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_avx2;
+    aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_avx2;
+    aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_avx2;
+    aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_avx2;
+    aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_avx2;
+    aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_avx2;
+    aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_avx2;
+    aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_avx2;
+    aom_highbd_sad_skip_4x16 = aom_highbd_sad_skip_4x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x16 = aom_highbd_sad_skip_4x16_sse2;
+    aom_highbd_sad_skip_4x16x4d = aom_highbd_sad_skip_4x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x16x4d = aom_highbd_sad_skip_4x16x4d_sse2;
+    aom_highbd_sad_skip_4x8 = aom_highbd_sad_skip_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x8 = aom_highbd_sad_skip_4x8_sse2;
+    aom_highbd_sad_skip_4x8x4d = aom_highbd_sad_skip_4x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_4x8x4d = aom_highbd_sad_skip_4x8x4d_sse2;
+    aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_avx2;
+    aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_avx2;
+    aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_avx2;
+    aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_avx2;
+    aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_avx2;
+    aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_avx2;
+    aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_avx2;
+    aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_avx2;
+    aom_highbd_sad_skip_8x16 = aom_highbd_sad_skip_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x16 = aom_highbd_sad_skip_8x16_sse2;
+    aom_highbd_sad_skip_8x16x4d = aom_highbd_sad_skip_8x16x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x16x4d = aom_highbd_sad_skip_8x16x4d_sse2;
+    aom_highbd_sad_skip_8x32 = aom_highbd_sad_skip_8x32_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x32 = aom_highbd_sad_skip_8x32_sse2;
+    aom_highbd_sad_skip_8x32x4d = aom_highbd_sad_skip_8x32x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x32x4d = aom_highbd_sad_skip_8x32x4d_sse2;
+    aom_highbd_sad_skip_8x8 = aom_highbd_sad_skip_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x8 = aom_highbd_sad_skip_8x8_sse2;
+    aom_highbd_sad_skip_8x8x4d = aom_highbd_sad_skip_8x8x4d_c;
+    if (flags & HAS_SSE2) aom_highbd_sad_skip_8x8x4d = aom_highbd_sad_skip_8x8x4d_sse2;
+    aom_highbd_sse = aom_highbd_sse_c;
+    if (flags & HAS_SSE4_1) aom_highbd_sse = aom_highbd_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_sse = aom_highbd_sse_avx2;
+    aom_highbd_subtract_block = aom_highbd_subtract_block_c;
+    if (flags & HAS_SSE2) aom_highbd_subtract_block = aom_highbd_subtract_block_sse2;
+    aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_sse2;
+    aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_sse2;
+    aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_sse2;
+    aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_sse2;
+    aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_sse2;
+    aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_sse2;
+    aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_sse2;
+    aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_sse2;
+    aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_sse2;
+    aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_sse2;
+    aom_ifft16x16_float = aom_ifft16x16_float_c;
+    if (flags & HAS_SSE2) aom_ifft16x16_float = aom_ifft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft16x16_float = aom_ifft16x16_float_avx2;
+    aom_ifft32x32_float = aom_ifft32x32_float_c;
+    if (flags & HAS_SSE2) aom_ifft32x32_float = aom_ifft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft32x32_float = aom_ifft32x32_float_avx2;
+    aom_ifft4x4_float = aom_ifft4x4_float_c;
+    if (flags & HAS_SSE2) aom_ifft4x4_float = aom_ifft4x4_float_sse2;
+    aom_ifft8x8_float = aom_ifft8x8_float_c;
+    if (flags & HAS_SSE2) aom_ifft8x8_float = aom_ifft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft8x8_float = aom_ifft8x8_float_avx2;
+    aom_int_pro_col = aom_int_pro_col_c;
+    if (flags & HAS_SSE2) aom_int_pro_col = aom_int_pro_col_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_col = aom_int_pro_col_avx2;
+    aom_int_pro_row = aom_int_pro_row_c;
+    if (flags & HAS_SSE2) aom_int_pro_row = aom_int_pro_row_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_row = aom_int_pro_row_avx2;
+    aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
+    aom_lpf_horizontal_14 = aom_lpf_horizontal_14_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_sse2;
+    aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_sse2;
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2;
+    aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_sse2;
+    aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_sse2;
+    aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_sse2;
+    aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_sse2;
+    aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_sse2;
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2;
+    aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_sse2;
+    aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_sse2;
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2;
+    aom_lpf_vertical_14 = aom_lpf_vertical_14_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_14 = aom_lpf_vertical_14_sse2;
+    aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_sse2;
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_avx2;
+    aom_lpf_vertical_4 = aom_lpf_vertical_4_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_4 = aom_lpf_vertical_4_sse2;
+    aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_sse2;
+    aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_sse2;
+    aom_lpf_vertical_6 = aom_lpf_vertical_6_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_6 = aom_lpf_vertical_6_sse2;
+    aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_sse2;
+    aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_sse2;
+    aom_lpf_vertical_8 = aom_lpf_vertical_8_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_8 = aom_lpf_vertical_8_sse2;
+    aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_sse2;
+    aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_sse2;
+    aom_masked_sad128x128 = aom_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128 = aom_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x128 = aom_masked_sad128x128_avx2;
+    aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_ssse3;
+    aom_masked_sad128x64 = aom_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64 = aom_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x64 = aom_masked_sad128x64_avx2;
+    aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_ssse3;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x16 = aom_masked_sad16x16_avx2;
+    aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_ssse3;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32 = aom_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x32 = aom_masked_sad16x32_avx2;
+    aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_ssse3;
+    aom_masked_sad16x4 = aom_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4 = aom_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x4 = aom_masked_sad16x4_avx2;
+    aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_ssse3;
+    aom_masked_sad16x64 = aom_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64 = aom_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x64 = aom_masked_sad16x64_avx2;
+    aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_ssse3;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8 = aom_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x8 = aom_masked_sad16x8_avx2;
+    aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_ssse3;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16 = aom_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x16 = aom_masked_sad32x16_avx2;
+    aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_ssse3;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32 = aom_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x32 = aom_masked_sad32x32_avx2;
+    aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_ssse3;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64 = aom_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x64 = aom_masked_sad32x64_avx2;
+    aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_ssse3;
+    aom_masked_sad32x8 = aom_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8 = aom_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x8 = aom_masked_sad32x8_avx2;
+    aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_ssse3;
+    aom_masked_sad4x16 = aom_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16 = aom_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x16 = aom_masked_sad4x16_avx2;
+    aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_ssse3;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4 = aom_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x4 = aom_masked_sad4x4_avx2;
+    aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_ssse3;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8 = aom_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x8 = aom_masked_sad4x8_avx2;
+    aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_ssse3;
+    aom_masked_sad64x128 = aom_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128 = aom_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x128 = aom_masked_sad64x128_avx2;
+    aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_ssse3;
+    aom_masked_sad64x16 = aom_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16 = aom_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x16 = aom_masked_sad64x16_avx2;
+    aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_ssse3;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32 = aom_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x32 = aom_masked_sad64x32_avx2;
+    aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_ssse3;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64 = aom_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x64 = aom_masked_sad64x64_avx2;
+    aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_ssse3;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16 = aom_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x16 = aom_masked_sad8x16_avx2;
+    aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_ssse3;
+    aom_masked_sad8x32 = aom_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32 = aom_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x32 = aom_masked_sad8x32_avx2;
+    aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_ssse3;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4 = aom_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x4 = aom_masked_sad8x4_avx2;
+    aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_ssse3;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8 = aom_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x8 = aom_masked_sad8x8_avx2;
+    aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_ssse3;
+    aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_ssse3;
+    aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_ssse3;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_ssse3;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_ssse3;
+    aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_ssse3;
+    aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_ssse3;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_ssse3;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_ssse3;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_ssse3;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_ssse3;
+    aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_ssse3;
+    aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_ssse3;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_ssse3;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_ssse3;
+    aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_ssse3;
+    aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_ssse3;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_ssse3;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_ssse3;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_ssse3;
+    aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_ssse3;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_ssse3;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_ssse3;
+    aom_minmax_8x8 = aom_minmax_8x8_c;
+    if (flags & HAS_SSE2) aom_minmax_8x8 = aom_minmax_8x8_sse2;
+    aom_mse16x16 = aom_mse16x16_c;
+    if (flags & HAS_SSE2) aom_mse16x16 = aom_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
+    aom_mse16x8 = aom_mse16x8_c;
+    if (flags & HAS_SSE2) aom_mse16x8 = aom_mse16x8_sse2;
+    aom_mse8x16 = aom_mse8x16_c;
+    if (flags & HAS_SSE2) aom_mse8x16 = aom_mse8x16_sse2;
+    aom_mse8x8 = aom_mse8x8_c;
+    if (flags & HAS_SSE2) aom_mse8x8 = aom_mse8x8_sse2;
+    aom_mse_16xh_16bit = aom_mse_16xh_16bit_c;
+    if (flags & HAS_SSE2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_avx2;
+    aom_mse_wxh_16bit = aom_mse_wxh_16bit_c;
+    if (flags & HAS_SSE2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_avx2;
+    aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_c;
+    if (flags & HAS_SSE2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_avx2;
+    aom_obmc_sad128x128 = aom_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x128 = aom_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x128 = aom_obmc_sad128x128_avx2;
+    aom_obmc_sad128x64 = aom_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x64 = aom_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x64 = aom_obmc_sad128x64_avx2;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x16 = aom_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x16 = aom_obmc_sad16x16_avx2;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x32 = aom_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x32 = aom_obmc_sad16x32_avx2;
+    aom_obmc_sad16x4 = aom_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x4 = aom_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x4 = aom_obmc_sad16x4_avx2;
+    aom_obmc_sad16x64 = aom_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x64 = aom_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x64 = aom_obmc_sad16x64_avx2;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x8 = aom_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x8 = aom_obmc_sad16x8_avx2;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x16 = aom_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x16 = aom_obmc_sad32x16_avx2;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x32 = aom_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x32 = aom_obmc_sad32x32_avx2;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x64 = aom_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x64 = aom_obmc_sad32x64_avx2;
+    aom_obmc_sad32x8 = aom_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x8 = aom_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x8 = aom_obmc_sad32x8_avx2;
+    aom_obmc_sad4x16 = aom_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x16 = aom_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x16 = aom_obmc_sad4x16_avx2;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x4 = aom_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x4 = aom_obmc_sad4x4_avx2;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x8 = aom_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x8 = aom_obmc_sad4x8_avx2;
+    aom_obmc_sad64x128 = aom_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x128 = aom_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x128 = aom_obmc_sad64x128_avx2;
+    aom_obmc_sad64x16 = aom_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x16 = aom_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x16 = aom_obmc_sad64x16_avx2;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x32 = aom_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x32 = aom_obmc_sad64x32_avx2;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x64 = aom_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x64 = aom_obmc_sad64x64_avx2;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x16 = aom_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x16 = aom_obmc_sad8x16_avx2;
+    aom_obmc_sad8x32 = aom_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x32 = aom_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x32 = aom_obmc_sad8x32_avx2;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x4 = aom_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x4 = aom_obmc_sad8x4_avx2;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x8 = aom_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x8 = aom_obmc_sad8x8_avx2;
+    aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_sse4_1;
+    aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_sse4_1;
+    aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_sse4_1;
+    aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_sse4_1;
+    aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_sse4_1;
+    aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_sse4_1;
+    aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_sse4_1;
+    aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_sse4_1;
+    aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_sse4_1;
+    aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_sse4_1;
+    aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_sse4_1;
+    aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_sse4_1;
+    aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_sse4_1;
+    aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_sse4_1;
+    aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_sse4_1;
+    aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_sse4_1;
+    aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_sse4_1;
+    aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_sse4_1;
+    aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_sse4_1;
+    aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_sse4_1;
+    aom_obmc_variance128x128 = aom_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x128 = aom_obmc_variance128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x128 = aom_obmc_variance128x128_avx2;
+    aom_obmc_variance128x64 = aom_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x64 = aom_obmc_variance128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x64 = aom_obmc_variance128x64_avx2;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x16 = aom_obmc_variance16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x16 = aom_obmc_variance16x16_avx2;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x32 = aom_obmc_variance16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x32 = aom_obmc_variance16x32_avx2;
+    aom_obmc_variance16x4 = aom_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x4 = aom_obmc_variance16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x4 = aom_obmc_variance16x4_avx2;
+    aom_obmc_variance16x64 = aom_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x64 = aom_obmc_variance16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x64 = aom_obmc_variance16x64_avx2;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x8 = aom_obmc_variance16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x8 = aom_obmc_variance16x8_avx2;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x16 = aom_obmc_variance32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x16 = aom_obmc_variance32x16_avx2;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x32 = aom_obmc_variance32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x32 = aom_obmc_variance32x32_avx2;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x64 = aom_obmc_variance32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x64 = aom_obmc_variance32x64_avx2;
+    aom_obmc_variance32x8 = aom_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x8 = aom_obmc_variance32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x8 = aom_obmc_variance32x8_avx2;
+    aom_obmc_variance4x16 = aom_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x16 = aom_obmc_variance4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x16 = aom_obmc_variance4x16_avx2;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x4 = aom_obmc_variance4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x4 = aom_obmc_variance4x4_avx2;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x8 = aom_obmc_variance4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x8 = aom_obmc_variance4x8_avx2;
+    aom_obmc_variance64x128 = aom_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x128 = aom_obmc_variance64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x128 = aom_obmc_variance64x128_avx2;
+    aom_obmc_variance64x16 = aom_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x16 = aom_obmc_variance64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x16 = aom_obmc_variance64x16_avx2;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x32 = aom_obmc_variance64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x32 = aom_obmc_variance64x32_avx2;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x64 = aom_obmc_variance64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x64 = aom_obmc_variance64x64_avx2;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x16 = aom_obmc_variance8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x16 = aom_obmc_variance8x16_avx2;
+    aom_obmc_variance8x32 = aom_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x32 = aom_obmc_variance8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x32 = aom_obmc_variance8x32_avx2;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x4 = aom_obmc_variance8x4_avx2;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x8 = aom_obmc_variance8x8_avx2;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_avx2;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_avx2;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_avx2;
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_avx2;
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_avx2;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3;
+    aom_quantize_b = aom_quantize_b_c;
+    if (flags & HAS_SSE2) aom_quantize_b = aom_quantize_b_sse2;
+    if (flags & HAS_AVX) aom_quantize_b = aom_quantize_b_avx;
+    if (flags & HAS_AVX2) aom_quantize_b = aom_quantize_b_avx2;
+    aom_quantize_b_32x32 = aom_quantize_b_32x32_c;
+    if (flags & HAS_AVX) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx;
+    if (flags & HAS_AVX2) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx2;
+    aom_quantize_b_32x32_adaptive = aom_quantize_b_32x32_adaptive_c;
+    if (flags & HAS_SSE2) aom_quantize_b_32x32_adaptive = aom_quantize_b_32x32_adaptive_sse2;
+    aom_quantize_b_64x64 = aom_quantize_b_64x64_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_64x64 = aom_quantize_b_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_quantize_b_64x64 = aom_quantize_b_64x64_avx2;
+    aom_quantize_b_64x64_adaptive = aom_quantize_b_64x64_adaptive_c;
+    if (flags & HAS_SSE2) aom_quantize_b_64x64_adaptive = aom_quantize_b_64x64_adaptive_sse2;
+    aom_quantize_b_adaptive = aom_quantize_b_adaptive_c;
+    if (flags & HAS_SSE2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_avx2;
+    aom_sad128x128 = aom_sad128x128_c;
+    if (flags & HAS_SSE2) aom_sad128x128 = aom_sad128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128 = aom_sad128x128_avx2;
+    aom_sad128x128_avg = aom_sad128x128_avg_c;
+    if (flags & HAS_SSE2) aom_sad128x128_avg = aom_sad128x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128_avg = aom_sad128x128_avg_avx2;
+    aom_sad128x128x3d = aom_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x128x3d = aom_sad128x128x3d_avx2;
+    aom_sad128x128x4d = aom_sad128x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad128x128x4d = aom_sad128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128x4d = aom_sad128x128x4d_avx2;
+    aom_sad128x64 = aom_sad128x64_c;
+    if (flags & HAS_SSE2) aom_sad128x64 = aom_sad128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64 = aom_sad128x64_avx2;
+    aom_sad128x64_avg = aom_sad128x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad128x64_avg = aom_sad128x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64_avg = aom_sad128x64_avg_avx2;
+    aom_sad128x64x3d = aom_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x64x3d = aom_sad128x64x3d_avx2;
+    aom_sad128x64x4d = aom_sad128x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad128x64x4d = aom_sad128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64x4d = aom_sad128x64x4d_avx2;
+    aom_sad16x16 = aom_sad16x16_c;
+    if (flags & HAS_SSE2) aom_sad16x16 = aom_sad16x16_sse2;
+    aom_sad16x16_avg = aom_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x16_avg = aom_sad16x16_avg_sse2;
+    aom_sad16x16x3d = aom_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x16x3d = aom_sad16x16x3d_avx2;
+    aom_sad16x16x4d = aom_sad16x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x16x4d = aom_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x16x4d = aom_sad16x16x4d_avx2;
+    aom_sad16x32 = aom_sad16x32_c;
+    if (flags & HAS_SSE2) aom_sad16x32 = aom_sad16x32_sse2;
+    aom_sad16x32_avg = aom_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x32_avg = aom_sad16x32_avg_sse2;
+    aom_sad16x32x3d = aom_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
+    aom_sad16x32x4d = aom_sad16x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x32x4d = aom_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4 = aom_sad16x4_c;
+    if (flags & HAS_SSE2) aom_sad16x4 = aom_sad16x4_sse2;
+    aom_sad16x4_avg = aom_sad16x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x4_avg = aom_sad16x4_avg_sse2;
+    aom_sad16x4x4d = aom_sad16x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x4x4d = aom_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
+    aom_sad16x64 = aom_sad16x64_c;
+    if (flags & HAS_SSE2) aom_sad16x64 = aom_sad16x64_sse2;
+    aom_sad16x64_avg = aom_sad16x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x64_avg = aom_sad16x64_avg_sse2;
+    aom_sad16x64x3d = aom_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x64x3d = aom_sad16x64x3d_avx2;
+    aom_sad16x64x4d = aom_sad16x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x64x4d = aom_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x64x4d = aom_sad16x64x4d_avx2;
+    aom_sad16x8 = aom_sad16x8_c;
+    if (flags & HAS_SSE2) aom_sad16x8 = aom_sad16x8_sse2;
+    aom_sad16x8_avg = aom_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x8_avg = aom_sad16x8_avg_sse2;
+    aom_sad16x8x3d = aom_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x8x3d = aom_sad16x8x3d_avx2;
+    aom_sad16x8x4d = aom_sad16x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x8x4d = aom_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x8x4d = aom_sad16x8x4d_avx2;
+    aom_sad32x16 = aom_sad32x16_c;
+    if (flags & HAS_SSE2) aom_sad32x16 = aom_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16 = aom_sad32x16_avx2;
+    aom_sad32x16_avg = aom_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x16_avg = aom_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16_avg = aom_sad32x16_avg_avx2;
+    aom_sad32x16x3d = aom_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x16x3d = aom_sad32x16x3d_avx2;
+    aom_sad32x16x4d = aom_sad32x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x16x4d = aom_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16x4d = aom_sad32x16x4d_avx2;
+    aom_sad32x32 = aom_sad32x32_c;
+    if (flags & HAS_SSE2) aom_sad32x32 = aom_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32 = aom_sad32x32_avx2;
+    aom_sad32x32_avg = aom_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x32_avg = aom_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32_avg = aom_sad32x32_avg_avx2;
+    aom_sad32x32x3d = aom_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x32x3d = aom_sad32x32x3d_avx2;
+    aom_sad32x32x4d = aom_sad32x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x32x4d = aom_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32x4d = aom_sad32x32x4d_avx2;
+    aom_sad32x64 = aom_sad32x64_c;
+    if (flags & HAS_SSE2) aom_sad32x64 = aom_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64 = aom_sad32x64_avx2;
+    aom_sad32x64_avg = aom_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x64_avg = aom_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64_avg = aom_sad32x64_avg_avx2;
+    aom_sad32x64x3d = aom_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x64x3d = aom_sad32x64x3d_avx2;
+    aom_sad32x64x4d = aom_sad32x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x64x4d = aom_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64x4d = aom_sad32x64x4d_avx2;
+    aom_sad32x8 = aom_sad32x8_c;
+    if (flags & HAS_SSE2) aom_sad32x8 = aom_sad32x8_sse2;
+    aom_sad32x8_avg = aom_sad32x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x8_avg = aom_sad32x8_avg_sse2;
+    aom_sad32x8x3d = aom_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x8x3d = aom_sad32x8x3d_avx2;
+    aom_sad32x8x4d = aom_sad32x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x8x4d = aom_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x8x4d = aom_sad32x8x4d_avx2;
+    aom_sad4x16 = aom_sad4x16_c;
+    if (flags & HAS_SSE2) aom_sad4x16 = aom_sad4x16_sse2;
+    aom_sad4x16_avg = aom_sad4x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x16_avg = aom_sad4x16_avg_sse2;
+    aom_sad4x16x4d = aom_sad4x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x16x4d = aom_sad4x16x4d_sse2;
+    aom_sad4x4 = aom_sad4x4_c;
+    if (flags & HAS_SSE2) aom_sad4x4 = aom_sad4x4_sse2;
+    aom_sad4x4_avg = aom_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x4_avg = aom_sad4x4_avg_sse2;
+    aom_sad4x4x4d = aom_sad4x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x4x4d = aom_sad4x4x4d_sse2;
+    aom_sad4x8 = aom_sad4x8_c;
+    if (flags & HAS_SSE2) aom_sad4x8 = aom_sad4x8_sse2;
+    aom_sad4x8_avg = aom_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x8_avg = aom_sad4x8_avg_sse2;
+    aom_sad4x8x4d = aom_sad4x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x8x4d = aom_sad4x8x4d_sse2;
+    aom_sad64x128 = aom_sad64x128_c;
+    if (flags & HAS_SSE2) aom_sad64x128 = aom_sad64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128 = aom_sad64x128_avx2;
+    aom_sad64x128_avg = aom_sad64x128_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x128_avg = aom_sad64x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128_avg = aom_sad64x128_avg_avx2;
+    aom_sad64x128x3d = aom_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x128x3d = aom_sad64x128x3d_avx2;
+    aom_sad64x128x4d = aom_sad64x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x128x4d = aom_sad64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128x4d = aom_sad64x128x4d_avx2;
+    aom_sad64x16 = aom_sad64x16_c;
+    if (flags & HAS_SSE2) aom_sad64x16 = aom_sad64x16_sse2;
+    aom_sad64x16_avg = aom_sad64x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x16_avg = aom_sad64x16_avg_sse2;
+    aom_sad64x16x3d = aom_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x16x3d = aom_sad64x16x3d_avx2;
+    aom_sad64x16x4d = aom_sad64x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x16x4d = aom_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x16x4d = aom_sad64x16x4d_avx2;
+    aom_sad64x32 = aom_sad64x32_c;
+    if (flags & HAS_SSE2) aom_sad64x32 = aom_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32 = aom_sad64x32_avx2;
+    aom_sad64x32_avg = aom_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x32_avg = aom_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32_avg = aom_sad64x32_avg_avx2;
+    aom_sad64x32x3d = aom_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x32x3d = aom_sad64x32x3d_avx2;
+    aom_sad64x32x4d = aom_sad64x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x32x4d = aom_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32x4d = aom_sad64x32x4d_avx2;
+    aom_sad64x64 = aom_sad64x64_c;
+    if (flags & HAS_SSE2) aom_sad64x64 = aom_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64 = aom_sad64x64_avx2;
+    aom_sad64x64_avg = aom_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x64_avg = aom_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64_avg = aom_sad64x64_avg_avx2;
+    aom_sad64x64x3d = aom_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x64x3d = aom_sad64x64x3d_avx2;
+    aom_sad64x64x4d = aom_sad64x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x64x4d = aom_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64x4d = aom_sad64x64x4d_avx2;
+    aom_sad8x16 = aom_sad8x16_c;
+    if (flags & HAS_SSE2) aom_sad8x16 = aom_sad8x16_sse2;
+    aom_sad8x16_avg = aom_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x16_avg = aom_sad8x16_avg_sse2;
+    aom_sad8x16x4d = aom_sad8x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x16x4d = aom_sad8x16x4d_sse2;
+    aom_sad8x32 = aom_sad8x32_c;
+    if (flags & HAS_SSE2) aom_sad8x32 = aom_sad8x32_sse2;
+    aom_sad8x32_avg = aom_sad8x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x32_avg = aom_sad8x32_avg_sse2;
+    aom_sad8x32x4d = aom_sad8x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x32x4d = aom_sad8x32x4d_sse2;
+    aom_sad8x4 = aom_sad8x4_c;
+    if (flags & HAS_SSE2) aom_sad8x4 = aom_sad8x4_sse2;
+    aom_sad8x4_avg = aom_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x4_avg = aom_sad8x4_avg_sse2;
+    aom_sad8x4x4d = aom_sad8x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x4x4d = aom_sad8x4x4d_sse2;
+    aom_sad8x8 = aom_sad8x8_c;
+    if (flags & HAS_SSE2) aom_sad8x8 = aom_sad8x8_sse2;
+    aom_sad8x8_avg = aom_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x8_avg = aom_sad8x8_avg_sse2;
+    aom_sad8x8x4d = aom_sad8x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x8x4d = aom_sad8x8x4d_sse2;
+    aom_sad_skip_128x128 = aom_sad_skip_128x128_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x128 = aom_sad_skip_128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128 = aom_sad_skip_128x128_avx2;
+    aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_avx2;
+    aom_sad_skip_128x64 = aom_sad_skip_128x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x64 = aom_sad_skip_128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64 = aom_sad_skip_128x64_avx2;
+    aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_avx2;
+    aom_sad_skip_16x16 = aom_sad_skip_16x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x16 = aom_sad_skip_16x16_sse2;
+    aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
+    aom_sad_skip_16x32 = aom_sad_skip_16x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x32 = aom_sad_skip_16x32_sse2;
+    aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x64 = aom_sad_skip_16x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x64 = aom_sad_skip_16x64_sse2;
+    aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
+    aom_sad_skip_16x8 = aom_sad_skip_16x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x8 = aom_sad_skip_16x8_sse2;
+    aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_avx2;
+    aom_sad_skip_32x16 = aom_sad_skip_32x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x16 = aom_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16 = aom_sad_skip_32x16_avx2;
+    aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_avx2;
+    aom_sad_skip_32x32 = aom_sad_skip_32x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x32 = aom_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32 = aom_sad_skip_32x32_avx2;
+    aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_avx2;
+    aom_sad_skip_32x64 = aom_sad_skip_32x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x64 = aom_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64 = aom_sad_skip_32x64_avx2;
+    aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_avx2;
+    aom_sad_skip_32x8 = aom_sad_skip_32x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x8 = aom_sad_skip_32x8_sse2;
+    aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_avx2;
+    aom_sad_skip_4x16 = aom_sad_skip_4x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x16 = aom_sad_skip_4x16_sse2;
+    aom_sad_skip_4x16x4d = aom_sad_skip_4x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x16x4d = aom_sad_skip_4x16x4d_sse2;
+    aom_sad_skip_4x8 = aom_sad_skip_4x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x8 = aom_sad_skip_4x8_sse2;
+    aom_sad_skip_4x8x4d = aom_sad_skip_4x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_4x8x4d = aom_sad_skip_4x8x4d_sse2;
+    aom_sad_skip_64x128 = aom_sad_skip_64x128_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x128 = aom_sad_skip_64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128 = aom_sad_skip_64x128_avx2;
+    aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_avx2;
+    aom_sad_skip_64x16 = aom_sad_skip_64x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x16 = aom_sad_skip_64x16_sse2;
+    aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_avx2;
+    aom_sad_skip_64x32 = aom_sad_skip_64x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x32 = aom_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32 = aom_sad_skip_64x32_avx2;
+    aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_avx2;
+    aom_sad_skip_64x64 = aom_sad_skip_64x64_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x64 = aom_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64 = aom_sad_skip_64x64_avx2;
+    aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_avx2;
+    aom_sad_skip_8x16 = aom_sad_skip_8x16_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x16 = aom_sad_skip_8x16_sse2;
+    aom_sad_skip_8x16x4d = aom_sad_skip_8x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x16x4d = aom_sad_skip_8x16x4d_sse2;
+    aom_sad_skip_8x32 = aom_sad_skip_8x32_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x32 = aom_sad_skip_8x32_sse2;
+    aom_sad_skip_8x32x4d = aom_sad_skip_8x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x32x4d = aom_sad_skip_8x32x4d_sse2;
+    aom_sad_skip_8x8 = aom_sad_skip_8x8_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x8 = aom_sad_skip_8x8_sse2;
+    aom_sad_skip_8x8x4d = aom_sad_skip_8x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad_skip_8x8x4d = aom_sad_skip_8x8x4d_sse2;
+    aom_satd = aom_satd_c;
+    if (flags & HAS_SSE2) aom_satd = aom_satd_sse2;
+    if (flags & HAS_AVX2) aom_satd = aom_satd_avx2;
+    aom_satd_lp = aom_satd_lp_c;
+    if (flags & HAS_SSE2) aom_satd_lp = aom_satd_lp_sse2;
+    if (flags & HAS_AVX2) aom_satd_lp = aom_satd_lp_avx2;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3;
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_ssse3;
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_ssse3;
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_ssse3;
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_ssse3;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_ssse3;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_ssse3;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_ssse3;
+    aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
+    aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
+    aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_ssse3;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_ssse3;
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_ssse3;
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_ssse3;
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_ssse3;
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_ssse3;
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_ssse3;
+    aom_sse = aom_sse_c;
+    if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_subtract_block = aom_subtract_block_c;
+    if (flags & HAS_SSE2) aom_subtract_block = aom_subtract_block_sse2;
+    if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_c;
+    if (flags & HAS_SSE2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_avx2;
+    aom_sum_squares_i16 = aom_sum_squares_i16_c;
+    if (flags & HAS_SSE2) aom_sum_squares_i16 = aom_sum_squares_i16_sse2;
+    aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_c;
+    if (flags & HAS_SSE2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_avx2;
+    aom_v_predictor_16x16 = aom_v_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x16 = aom_v_predictor_16x16_sse2;
+    aom_v_predictor_16x32 = aom_v_predictor_16x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x32 = aom_v_predictor_16x32_sse2;
+    aom_v_predictor_16x4 = aom_v_predictor_16x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x4 = aom_v_predictor_16x4_sse2;
+    aom_v_predictor_16x64 = aom_v_predictor_16x64_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x64 = aom_v_predictor_16x64_sse2;
+    aom_v_predictor_16x8 = aom_v_predictor_16x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x8 = aom_v_predictor_16x8_sse2;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x16 = aom_v_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x16 = aom_v_predictor_32x16_avx2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x32 = aom_v_predictor_32x32_avx2;
+    aom_v_predictor_32x64 = aom_v_predictor_32x64_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+    aom_v_predictor_32x8 = aom_v_predictor_32x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x8 = aom_v_predictor_32x8_sse2;
+    aom_v_predictor_4x16 = aom_v_predictor_4x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x16 = aom_v_predictor_4x16_sse2;
+    aom_v_predictor_4x4 = aom_v_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x4 = aom_v_predictor_4x4_sse2;
+    aom_v_predictor_4x8 = aom_v_predictor_4x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x8 = aom_v_predictor_4x8_sse2;
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
+    aom_v_predictor_64x32 = aom_v_predictor_64x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
+    aom_v_predictor_64x64 = aom_v_predictor_64x64_c;
+    if (flags & HAS_SSE2) aom_v_predictor_64x64 = aom_v_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x64 = aom_v_predictor_64x64_avx2;
+    aom_v_predictor_8x16 = aom_v_predictor_8x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x16 = aom_v_predictor_8x16_sse2;
+    aom_v_predictor_8x32 = aom_v_predictor_8x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x32 = aom_v_predictor_8x32_sse2;
+    aom_v_predictor_8x4 = aom_v_predictor_8x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x4 = aom_v_predictor_8x4_sse2;
+    aom_v_predictor_8x8 = aom_v_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x8 = aom_v_predictor_8x8_sse2;
+    aom_var_2d_u16 = aom_var_2d_u16_c;
+    if (flags & HAS_SSE2) aom_var_2d_u16 = aom_var_2d_u16_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u16 = aom_var_2d_u16_avx2;
+    aom_var_2d_u8 = aom_var_2d_u8_c;
+    if (flags & HAS_SSE2) aom_var_2d_u8 = aom_var_2d_u8_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u8 = aom_var_2d_u8_avx2;
+    aom_variance128x128 = aom_variance128x128_c;
+    if (flags & HAS_SSE2) aom_variance128x128 = aom_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_variance128x128 = aom_variance128x128_avx2;
+    aom_variance128x64 = aom_variance128x64_c;
+    if (flags & HAS_SSE2) aom_variance128x64 = aom_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_variance128x64 = aom_variance128x64_avx2;
+    aom_variance16x16 = aom_variance16x16_c;
+    if (flags & HAS_SSE2) aom_variance16x16 = aom_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
+    aom_variance16x32 = aom_variance16x32_c;
+    if (flags & HAS_SSE2) aom_variance16x32 = aom_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_variance16x32 = aom_variance16x32_avx2;
+    aom_variance16x4 = aom_variance16x4_c;
+    if (flags & HAS_SSE2) aom_variance16x4 = aom_variance16x4_sse2;
+    if (flags & HAS_AVX2) aom_variance16x4 = aom_variance16x4_avx2;
+    aom_variance16x64 = aom_variance16x64_c;
+    if (flags & HAS_SSE2) aom_variance16x64 = aom_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_variance16x64 = aom_variance16x64_avx2;
+    aom_variance16x8 = aom_variance16x8_c;
+    if (flags & HAS_SSE2) aom_variance16x8 = aom_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_variance16x8 = aom_variance16x8_avx2;
+    aom_variance32x16 = aom_variance32x16_c;
+    if (flags & HAS_SSE2) aom_variance32x16 = aom_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_variance32x16 = aom_variance32x16_avx2;
+    aom_variance32x32 = aom_variance32x32_c;
+    if (flags & HAS_SSE2) aom_variance32x32 = aom_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_variance32x32 = aom_variance32x32_avx2;
+    aom_variance32x64 = aom_variance32x64_c;
+    if (flags & HAS_SSE2) aom_variance32x64 = aom_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_variance32x64 = aom_variance32x64_avx2;
+    aom_variance32x8 = aom_variance32x8_c;
+    if (flags & HAS_SSE2) aom_variance32x8 = aom_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_variance32x8 = aom_variance32x8_avx2;
+    aom_variance4x16 = aom_variance4x16_c;
+    if (flags & HAS_SSE2) aom_variance4x16 = aom_variance4x16_sse2;
+    aom_variance4x4 = aom_variance4x4_c;
+    if (flags & HAS_SSE2) aom_variance4x4 = aom_variance4x4_sse2;
+    aom_variance4x8 = aom_variance4x8_c;
+    if (flags & HAS_SSE2) aom_variance4x8 = aom_variance4x8_sse2;
+    aom_variance64x128 = aom_variance64x128_c;
+    if (flags & HAS_SSE2) aom_variance64x128 = aom_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_variance64x128 = aom_variance64x128_avx2;
+    aom_variance64x16 = aom_variance64x16_c;
+    if (flags & HAS_SSE2) aom_variance64x16 = aom_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_variance64x16 = aom_variance64x16_avx2;
+    aom_variance64x32 = aom_variance64x32_c;
+    if (flags & HAS_SSE2) aom_variance64x32 = aom_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_variance64x32 = aom_variance64x32_avx2;
+    aom_variance64x64 = aom_variance64x64_c;
+    if (flags & HAS_SSE2) aom_variance64x64 = aom_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_variance64x64 = aom_variance64x64_avx2;
+    aom_variance8x16 = aom_variance8x16_c;
+    if (flags & HAS_SSE2) aom_variance8x16 = aom_variance8x16_sse2;
+    aom_variance8x32 = aom_variance8x32_c;
+    if (flags & HAS_SSE2) aom_variance8x32 = aom_variance8x32_sse2;
+    aom_variance8x4 = aom_variance8x4_c;
+    if (flags & HAS_SSE2) aom_variance8x4 = aom_variance8x4_sse2;
+    aom_variance8x8 = aom_variance8x8_c;
+    if (flags & HAS_SSE2) aom_variance8x8 = aom_variance8x8_sse2;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
+    if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
+    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
+    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/win/ia32/config/aom_scale_rtcd.h b/media/libaom/config/win/ia32/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..3b70fb47c3
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_scale_rtcd.h
@@ -0,0 +1,103 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/win/ia32/config/av1_rtcd.h b/media/libaom/config/win/ia32/config/av1_rtcd.h
new file mode 100644
index 0000000000..3f404f61c8
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/av1_rtcd.h
@@ -0,0 +1,1272 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+void aom_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+RTCD_EXTERN void (*aom_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_dist_wtd_comp_avg_upsampled_pred_ssse3(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+RTCD_EXTERN void (*aom_highbd_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+#define aom_quantize_b_helper aom_quantize_b_helper_c
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+RTCD_EXTERN void (*aom_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_sse4_1(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_avx2(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_avx2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+RTCD_EXTERN int64_t (*av1_block_error_lp)(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim1)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim2)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params_high_bd)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+void av1_cnn_convolve_no_maxpool_padding_valid_avx2(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+RTCD_EXTERN void (*av1_cnn_convolve_no_maxpool_padding_valid)(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+RTCD_EXTERN void (*av1_compute_stats)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+RTCD_EXTERN void (*av1_compute_stats_highbd)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr_intrabc av1_convolve_2d_sr_intrabc_c
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr_intrabc av1_convolve_x_sr_intrabc_c
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+double av1_estimate_noise_from_single_plane_avx2(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+RTCD_EXTERN double (*av1_estimate_noise_from_single_plane)(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void av1_fwht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*av1_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+uint32_t av1_get_crc32c_value_sse4_2(void *crc_calculator, uint8_t *p, size_t length);
+RTCD_EXTERN uint32_t (*av1_get_crc32c_value)(void *crc_calculator, uint8_t *p, size_t length);
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_sse4_1(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_avx2(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+RTCD_EXTERN void (*av1_get_horver_correlation_full)(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+void av1_get_nz_map_contexts_sse2(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+RTCD_EXTERN void (*av1_get_nz_map_contexts)(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_highbd_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+RTCD_EXTERN int64_t (*av1_highbd_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_c
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_c
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr_intrabc av1_highbd_convolve_2d_sr_intrabc_c
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr_intrabc av1_highbd_convolve_x_sr_intrabc_c
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr_intrabc av1_highbd_convolve_y_sr_intrabc_c
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+#define av1_highbd_estimate_noise_from_single_plane av1_highbd_estimate_noise_from_single_plane_c
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+void av1_highbd_filter_intra_edge_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_highbd_filter_intra_edge)(uint16_t *p, int sz, int strength);
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_highbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_sse4_1(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+RTCD_EXTERN void (*av1_highbd_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+void av1_highbd_upsample_intra_edge_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_highbd_upsample_intra_edge)(uint16_t *p, int sz, int bd);
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_lowbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+void av1_nn_fast_softmax_16_sse3(const float *input_nodes, float *output);
+RTCD_EXTERN void (*av1_nn_fast_softmax_16)(const float *input_nodes, float *output);
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_sse3(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_avx2(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+RTCD_EXTERN void (*av1_nn_predict)(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_avx2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_lp)(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit);
+RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit);
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+RTCD_EXTERN void (*av1_txb_init_levels)(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_avx2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+RTCD_EXTERN void (*av1_wedge_compute_delta_squares)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_avx2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+RTCD_EXTERN int8_t (*av1_wedge_sign_from_residuals)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+RTCD_EXTERN uint64_t (*av1_wedge_sse_from_residuals)(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*cfl_get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*cfl_get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*cfl_get_subtract_average_fn)(TX_SIZE tx_size);
+
+void av1_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_sse2;
+    aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+    aom_highbd_comp_avg_upsampled_pred = aom_highbd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_comp_avg_upsampled_pred = aom_highbd_comp_avg_upsampled_pred_sse2;
+    aom_highbd_dist_wtd_comp_avg_upsampled_pred = aom_highbd_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_dist_wtd_comp_avg_upsampled_pred = aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2;
+    aom_highbd_upsampled_pred = aom_highbd_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_highbd_upsampled_pred = aom_highbd_upsampled_pred_sse2;
+    aom_upsampled_pred = aom_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_upsampled_pred = aom_upsampled_pred_sse2;
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
+    av1_apply_temporal_filter = av1_apply_temporal_filter_c;
+    if (flags & HAS_SSE2) av1_apply_temporal_filter = av1_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_apply_temporal_filter = av1_apply_temporal_filter_avx2;
+    av1_block_error = av1_block_error_c;
+    if (flags & HAS_SSE2) av1_block_error = av1_block_error_sse2;
+    if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_block_error_lp = av1_block_error_lp_c;
+    if (flags & HAS_SSE2) av1_block_error_lp = av1_block_error_lp_sse2;
+    if (flags & HAS_AVX2) av1_block_error_lp = av1_block_error_lp_avx2;
+    av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
+    av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
+    av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
+    if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
+    av1_calc_indices_dim1 = av1_calc_indices_dim1_c;
+    if (flags & HAS_SSE2) av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
+    av1_calc_indices_dim2 = av1_calc_indices_dim2_c;
+    if (flags & HAS_SSE2) av1_calc_indices_dim2 = av1_calc_indices_dim2_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim2 = av1_calc_indices_dim2_avx2;
+    av1_calc_proj_params = av1_calc_proj_params_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params = av1_calc_proj_params_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params = av1_calc_proj_params_avx2;
+    av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_avx2;
+    av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_c;
+    if (flags & HAS_AVX2) av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_avx2;
+    av1_compute_stats = av1_compute_stats_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats = av1_compute_stats_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats = av1_compute_stats_avx2;
+    av1_compute_stats_highbd = av1_compute_stats_highbd_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats_highbd = av1_compute_stats_highbd_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats_highbd = av1_compute_stats_highbd_avx2;
+    av1_convolve_2d_scale = av1_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
+    av1_convolve_2d_sr = av1_convolve_2d_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
+    av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
+    av1_convolve_x_sr = av1_convolve_x_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_x_sr = av1_convolve_x_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
+    av1_convolve_y_sr = av1_convolve_y_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_y_sr = av1_convolve_y_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
+    av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_avx2;
+    av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_avx2;
+    av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2;
+    av1_dr_prediction_z1 = av1_dr_prediction_z1_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2;
+    av1_dr_prediction_z2 = av1_dr_prediction_z2_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2;
+    av1_dr_prediction_z3 = av1_dr_prediction_z3_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2;
+    av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_c;
+    if (flags & HAS_AVX2) av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_avx2;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
+    av1_filter_intra_predictor = av1_filter_intra_predictor_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_avx2;
+    av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_sse4_1;
+    av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_sse4_1;
+    av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_sse4_1;
+    av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_avx2;
+    av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_sse4_1;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_avx2;
+    av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_sse4_1;
+    av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_sse4_1;
+    av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_sse4_1;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
+    av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_sse4_1;
+    av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_sse4_1;
+    av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_sse4_1;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_avx2;
+    av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_avx2;
+    av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_sse4_1;
+    av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_sse4_1;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_avx2;
+    av1_fwht4x4 = av1_fwht4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwht4x4 = av1_fwht4x4_sse4_1;
+    av1_get_crc32c_value = av1_get_crc32c_value_c;
+    if (flags & HAS_SSE4_2) av1_get_crc32c_value = av1_get_crc32c_value_sse4_2;
+    av1_get_horver_correlation_full = av1_get_horver_correlation_full_c;
+    if (flags & HAS_SSE4_1) av1_get_horver_correlation_full = av1_get_horver_correlation_full_sse4_1;
+    if (flags & HAS_AVX2) av1_get_horver_correlation_full = av1_get_horver_correlation_full_avx2;
+    av1_get_nz_map_contexts = av1_get_nz_map_contexts_c;
+    if (flags & HAS_SSE2) av1_get_nz_map_contexts = av1_get_nz_map_contexts_sse2;
+    av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_c;
+    if (flags & HAS_SSE2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_avx2;
+    av1_highbd_block_error = av1_highbd_block_error_c;
+    if (flags & HAS_SSE2) av1_highbd_block_error = av1_highbd_block_error_sse2;
+    if (flags & HAS_AVX2) av1_highbd_block_error = av1_highbd_block_error_avx2;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
+    av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
+    av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
+    av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
+    av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_avx2;
+    av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_avx2;
+    av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_avx2;
+    av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_avx2;
+    av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_avx2;
+    av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_avx2;
+    av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_avx2;
+    av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_sse4_1;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_sse4_1;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_sse4_1;
+    av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
+    av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1;
+    av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_avx2;
+    av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
+    if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
+    av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_sse4_1;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2;
+    av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
+    if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
+    av1_inv_txfm_add = av1_inv_txfm_add_c;
+    if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
+    if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_c;
+    if (flags & HAS_SSE2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
+    av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_avx2;
+    av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_c;
+    if (flags & HAS_SSE3) av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_sse3;
+    av1_nn_predict = av1_nn_predict_c;
+    if (flags & HAS_SSE3) av1_nn_predict = av1_nn_predict_sse3;
+    if (flags & HAS_AVX2) av1_nn_predict = av1_nn_predict_avx2;
+    av1_quantize_fp = av1_quantize_fp_c;
+    if (flags & HAS_SSE2) av1_quantize_fp = av1_quantize_fp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
+    av1_quantize_fp_64x64 = av1_quantize_fp_64x64_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_64x64 = av1_quantize_fp_64x64_avx2;
+    av1_quantize_lp = av1_quantize_lp_c;
+    if (flags & HAS_SSE2) av1_quantize_lp = av1_quantize_lp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_lp = av1_quantize_lp_avx2;
+    av1_resize_and_extend_frame = av1_resize_and_extend_frame_c;
+    if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
+    av1_txb_init_levels = av1_txb_init_levels_c;
+    if (flags & HAS_SSE4_1) av1_txb_init_levels = av1_txb_init_levels_sse4_1;
+    if (flags & HAS_AVX2) av1_txb_init_levels = av1_txb_init_levels_avx2;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_warp_affine = av1_warp_affine_avx2;
+    av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_c;
+    if (flags & HAS_SSE2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
+    if (flags & HAS_AVX2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
+    av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_c;
+    if (flags & HAS_SSE2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_avx2;
+    av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_c;
+    if (flags & HAS_SSE2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
+    av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_c;
+    if (flags & HAS_SSE2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
+    if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
+    if (flags & HAS_SSE2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
+    if (flags & HAS_SSE2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
+    cdef_filter_16_0 = cdef_filter_16_0_c;
+    if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
+    cdef_filter_16_1 = cdef_filter_16_1_c;
+    if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
+    cdef_filter_16_2 = cdef_filter_16_2_c;
+    if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
+    cdef_filter_16_3 = cdef_filter_16_3_c;
+    if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
+    cdef_filter_8_0 = cdef_filter_8_0_c;
+    if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
+    cdef_filter_8_1 = cdef_filter_8_1_c;
+    if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
+    cdef_filter_8_2 = cdef_filter_8_2_c;
+    if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
+    cdef_filter_8_3 = cdef_filter_8_3_c;
+    if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
+    cdef_find_dir = cdef_find_dir_c;
+    if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
+    cdef_find_dir_dual = cdef_find_dir_dual_c;
+    if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
+    cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
+    cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
+    cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
+    cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
+    cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
+    cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
+    cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_avx2;
+    cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_avx2;
+    cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_c;
+    if (flags & HAS_SSE2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/win/x64/config/aom_config.asm b/media/libaom/config/win/x64/config/aom_config.asm
new file mode 100644
index 0000000000..f793ff3c6d
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_config.asm
@@ -0,0 +1,92 @@
+;
+; Copyright (c) 2024, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+AOM_ARCH_AARCH64 equ 0
+AOM_ARCH_ARM equ 0
+AOM_ARCH_PPC equ 0
+AOM_ARCH_X86 equ 0
+AOM_ARCH_X86_64 equ 1
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_HIGHBITDEPTH equ 1
+CONFIG_AV1_TEMPORAL_DENOISING equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITRATE_ACCURACY equ 0
+CONFIG_BITRATE_ACCURACY_BL equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_COMPONENT_TIMING equ 0
+CONFIG_COLLECT_PARTITION_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_CWG_C013 equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_EXCLUDE_SIMD_MISMATCH equ 0
+CONFIG_FPMT_TEST equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBVMAF_PSNR_PEAK equ 1
+CONFIG_LIBYUV equ 0
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NN_V2 equ 0
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OPTICAL_FLOW_API equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_OUTPUT_FRAME_SIZE equ 0
+CONFIG_PARTITION_SEARCH_ORDER equ 0
+CONFIG_PIC equ 0
+CONFIG_RATECTRL_LOG equ 0
+CONFIG_RD_COMMAND equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_RT_ML_PARTITIONING equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SALIENCY_MAP equ 0
+CONFIG_SHARED equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_SPEED_STATS equ 0
+CONFIG_TFLITE equ 0
+CONFIG_THREE_PASS equ 0
+CONFIG_TUNE_BUTTERAUGLI equ 0
+CONFIG_TUNE_VMAF equ 0
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+FORCE_HIGHBITDEPTH_DECODING equ 0
+HAVE_ARM_CRC32 equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_FEXCEPT equ 1
+HAVE_MMX equ 1
+HAVE_NEON equ 0
+HAVE_NEON_DOTPROD equ 0
+HAVE_NEON_I8MM equ 0
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 1
+HAVE_SSE4_1 equ 1
+HAVE_SSE4_2 equ 1
+HAVE_SSSE3 equ 1
+HAVE_SVE equ 0
+HAVE_VSX equ 0
+HAVE_WXWIDGETS equ 0
+STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/win/x64/config/aom_config.h b/media/libaom/config/win/x64/config/aom_config.h
new file mode 100644
index 0000000000..670d2ffe56
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_config.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2024, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_CONFIG_H_
+#define AOM_CONFIG_H_
+
+#define AOM_ARCH_AARCH64 0
+#define AOM_ARCH_ARM 0
+#define AOM_ARCH_PPC 0
+#define AOM_ARCH_X86 0
+#define AOM_ARCH_X86_64 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_HIGHBITDEPTH 1
+#define CONFIG_AV1_TEMPORAL_DENOISING 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITRATE_ACCURACY 0
+#define CONFIG_BITRATE_ACCURACY_BL 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_COMPONENT_TIMING 0
+#define CONFIG_COLLECT_PARTITION_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_CWG_C013 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_EXCLUDE_SIMD_MISMATCH 0
+#define CONFIG_FPMT_TEST 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBVMAF_PSNR_PEAK 1
+#define CONFIG_LIBYUV 0
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NN_V2 0
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OPTICAL_FLOW_API 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_OUTPUT_FRAME_SIZE 0
+#define CONFIG_PARTITION_SEARCH_ORDER 0
+#define CONFIG_PIC 0
+#define CONFIG_RATECTRL_LOG 0
+#define CONFIG_RD_COMMAND 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_RT_ML_PARTITIONING 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SALIENCY_MAP 0
+#define CONFIG_SHARED 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_SPEED_STATS 0
+#define CONFIG_TFLITE 0
+#define CONFIG_THREE_PASS 0
+#define CONFIG_TUNE_BUTTERAUGLI 0
+#define CONFIG_TUNE_VMAF 0
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define FORCE_HIGHBITDEPTH_DECODING 0
+#define HAVE_ARM_CRC32 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_FEXCEPT 1
+#define HAVE_MMX 1
+#define HAVE_NEON 0
+#define HAVE_NEON_DOTPROD 0
+#define HAVE_NEON_I8MM 0
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_SSE4_2 1
+#define HAVE_SSSE3 1
+#define HAVE_SVE 0
+#define HAVE_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INLINE inline
+#define STATIC_LINK_JXL 0
+#endif /* AOM_CONFIG_H_ */
diff --git a/media/libaom/config/win/x64/config/aom_dsp_rtcd.h b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h
new file mode 100644
index 0000000000..8e979cc189
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h
@@ -0,0 +1,8036 @@
+// This file is generated. Do not edit.
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+#define aom_avg_4x4 aom_avg_4x4_sse2
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+#define aom_avg_8x8 aom_avg_8x8_sse2
+
+void aom_avg_8x8_quad_c(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+void aom_avg_8x8_quad_avx2(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+RTCD_EXTERN void (*aom_avg_8x8_quad)(const uint8_t *s, int p, int x16_idx, int y16_idx, int *avg);
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh);
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_sse2
+
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
+
+void aom_dc_128_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x4 aom_dc_128_predictor_16x4_sse2
+
+void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x64 aom_dc_128_predictor_16x64_sse2
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
+
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x8 aom_dc_128_predictor_32x8_sse2
+
+void aom_dc_128_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x16 aom_dc_128_predictor_4x16_sse2
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_sse2
+
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
+
+void aom_dc_128_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
+
+void aom_dc_128_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x32 aom_dc_128_predictor_8x32_sse2
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_sse2
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_sse2
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_sse2
+
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
+
+void aom_dc_left_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x4 aom_dc_left_predictor_16x4_sse2
+
+void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x64 aom_dc_left_predictor_16x64_sse2
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
+
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x8 aom_dc_left_predictor_32x8_sse2
+
+void aom_dc_left_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x16 aom_dc_left_predictor_4x16_sse2
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_sse2
+
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
+
+void aom_dc_left_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
+
+void aom_dc_left_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x32 aom_dc_left_predictor_8x32_sse2
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_sse2
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_sse2
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x16 aom_dc_predictor_16x16_sse2
+
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
+
+void aom_dc_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x4 aom_dc_predictor_16x4_sse2
+
+void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x64 aom_dc_predictor_16x64_sse2
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
+
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x8 aom_dc_predictor_32x8_sse2
+
+void aom_dc_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x16 aom_dc_predictor_4x16_sse2
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x4 aom_dc_predictor_4x4_sse2
+
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
+
+void aom_dc_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
+
+void aom_dc_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x32 aom_dc_predictor_8x32_sse2
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_sse2
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x8 aom_dc_predictor_8x8_sse2
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_sse2
+
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
+
+void aom_dc_top_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x4 aom_dc_top_predictor_16x4_sse2
+
+void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x64 aom_dc_top_predictor_16x64_sse2
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
+
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x8 aom_dc_top_predictor_32x8_sse2
+
+void aom_dc_top_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x16 aom_dc_top_predictor_4x16_sse2
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_sse2
+
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
+
+void aom_dc_top_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
+
+void aom_dc_top_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x32 aom_dc_top_predictor_8x32_sse2
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_sse2
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_sse2
+
+void aom_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_dist_wtd_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+unsigned int aom_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x128_avg aom_dist_wtd_sad128x128_avg_sse2
+
+unsigned int aom_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad128x64_avg aom_dist_wtd_sad128x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x16_avg aom_dist_wtd_sad16x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x32_avg aom_dist_wtd_sad16x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x4_avg aom_dist_wtd_sad16x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x64_avg aom_dist_wtd_sad16x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad16x8_avg aom_dist_wtd_sad16x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x16_avg aom_dist_wtd_sad32x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x32_avg aom_dist_wtd_sad32x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x64_avg aom_dist_wtd_sad32x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad32x8_avg aom_dist_wtd_sad32x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x16_avg aom_dist_wtd_sad4x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x4_avg aom_dist_wtd_sad4x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad4x8_avg aom_dist_wtd_sad4x8_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x128_avg aom_dist_wtd_sad64x128_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x16_avg aom_dist_wtd_sad64x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x32_avg aom_dist_wtd_sad64x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad64x64_avg aom_dist_wtd_sad64x64_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x16_avg aom_dist_wtd_sad8x16_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x32_avg aom_dist_wtd_sad8x32_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x4_avg aom_dist_wtd_sad8x4_avg_sse2
+
+unsigned int aom_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+unsigned int aom_dist_wtd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_dist_wtd_sad8x8_avg aom_dist_wtd_sad8x8_avg_sse2
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+uint32_t aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+RTCD_EXTERN uint32_t (*aom_dist_wtd_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_sse2
+
+void aom_fdct4x4_lp_c(const int16_t *input, int16_t *output, int stride);
+void aom_fdct4x4_lp_sse2(const int16_t *input, int16_t *output, int stride);
+#define aom_fdct4x4_lp aom_fdct4x4_lp_sse2
+
+void aom_fft16x16_float_c(const float *input, float *temp, float *output);
+void aom_fft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_fft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_fft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_fft2x2_float aom_fft2x2_float_c
+
+void aom_fft32x32_float_c(const float *input, float *temp, float *output);
+void aom_fft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_fft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_fft4x4_float_c(const float *input, float *temp, float *output);
+void aom_fft4x4_float_sse2(const float *input, float *temp, float *output);
+#define aom_fft4x4_float aom_fft4x4_float_sse2
+
+void aom_fft8x8_float_c(const float *input, float *temp, float *output);
+void aom_fft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_fft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_fft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_get_blk_sse_sum_c(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_sse2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+void aom_get_blk_sse_sum_avx2(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+RTCD_EXTERN void (*aom_get_blk_sse_sum)(const int16_t *data, int stride, int bw, int bh, int *x_sum, int64_t *x2_sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_sse2
+
+void aom_get_var_sse_sum_16x16_dual_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+void aom_get_var_sse_sum_16x16_dual_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+RTCD_EXTERN void (*aom_get_var_sse_sum_16x16_dual)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse16x16, unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16);
+
+void aom_get_var_sse_sum_8x8_quad_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+void aom_get_var_sse_sum_8x8_quad_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+RTCD_EXTERN void (*aom_get_var_sse_sum_8x8_quad)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse8x8, int *sum8x8, unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x16 aom_h_predictor_16x16_sse2
+
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
+
+void aom_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x4 aom_h_predictor_16x4_sse2
+
+void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x64 aom_h_predictor_16x64_sse2
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
+
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2
+
+void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x64 aom_h_predictor_32x64_sse2
+
+void aom_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x8 aom_h_predictor_32x8_sse2
+
+void aom_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x16 aom_h_predictor_4x16_sse2
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x4 aom_h_predictor_4x4_sse2
+
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
+
+void aom_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x16 aom_h_predictor_64x16_sse2
+
+void aom_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x32 aom_h_predictor_64x32_sse2
+
+void aom_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_64x64 aom_h_predictor_64x64_sse2
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
+
+void aom_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x32 aom_h_predictor_8x32_sse2
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_sse2
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x8 aom_h_predictor_8x8_sse2
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_hadamard_4x4_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_4x4 aom_hadamard_4x4_sse2
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+#define aom_hadamard_8x8 aom_hadamard_8x8_sse2
+
+void aom_hadamard_lp_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+void aom_hadamard_lp_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+#define aom_hadamard_lp_8x8 aom_hadamard_lp_8x8_sse2
+
+void aom_hadamard_lp_8x8_dual_c(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+void aom_hadamard_lp_8x8_dual_avx2(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_lp_8x8_dual)(const int16_t *src_diff, ptrdiff_t src_stride, int16_t *coeff);
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_10_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_10_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_sse2
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x128 aom_highbd_10_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance128x64 aom_highbd_10_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x64 aom_highbd_10_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x128 aom_highbd_10_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x16 aom_highbd_10_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x128 aom_highbd_10_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance128x64 aom_highbd_10_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x4 aom_highbd_10_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x64 aom_highbd_10_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x8 aom_highbd_10_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x128 aom_highbd_10_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x16 aom_highbd_10_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x32 aom_highbd_10_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x4 aom_highbd_10_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x64 aom_highbd_10_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x8 aom_highbd_10_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x16 aom_highbd_10_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x32 aom_highbd_10_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
+unsigned int aom_highbd_10_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_10_variance8x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_12_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_12_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_sse2
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_sse2
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x128 aom_highbd_12_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance128x64 aom_highbd_12_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x64 aom_highbd_12_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x128 aom_highbd_12_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x16 aom_highbd_12_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x128 aom_highbd_12_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance128x64 aom_highbd_12_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x4 aom_highbd_12_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x64 aom_highbd_12_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x8 aom_highbd_12_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x128 aom_highbd_12_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x16 aom_highbd_12_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x32 aom_highbd_12_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x128 aom_highbd_12_sub_pixel_variance128x128_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance128x64 aom_highbd_12_sub_pixel_variance128x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x4 aom_highbd_12_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x64 aom_highbd_12_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x8 aom_highbd_12_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x128 aom_highbd_12_sub_pixel_variance64x128_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x16 aom_highbd_12_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x32 aom_highbd_12_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_12_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x128 aom_highbd_12_variance128x128_sse2
+
+unsigned int aom_highbd_12_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance128x64 aom_highbd_12_variance128x64_sse2
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_sse2
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_sse2
+
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
+unsigned int aom_highbd_12_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x64 aom_highbd_12_variance16x64_sse2
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_sse2
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_sse2
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_sse2
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_sse2
+
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x8 aom_highbd_12_variance32x8_sse2
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x128 aom_highbd_12_variance64x128_sse2
+
+unsigned int aom_highbd_12_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x16 aom_highbd_12_variance64x16_sse2
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_sse2
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_sse2
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_sse2
+
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x32 aom_highbd_12_variance8x32_sse2
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_12_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_sse2
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64 aom_highbd_8_dist_wtd_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x32_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8 aom_highbd_8_dist_wtd_sub_pixel_avg_variance8x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_sse2
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_sse2
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x128 aom_highbd_8_obmc_sub_pixel_variance128x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance128x64 aom_highbd_8_obmc_sub_pixel_variance128x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x16 aom_highbd_8_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x32 aom_highbd_8_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x4 aom_highbd_8_obmc_sub_pixel_variance16x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x64 aom_highbd_8_obmc_sub_pixel_variance16x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance16x8 aom_highbd_8_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x16 aom_highbd_8_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x32 aom_highbd_8_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x64 aom_highbd_8_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance32x8 aom_highbd_8_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x16 aom_highbd_8_obmc_sub_pixel_variance4x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x4 aom_highbd_8_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance4x8 aom_highbd_8_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x128 aom_highbd_8_obmc_sub_pixel_variance64x128_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x16 aom_highbd_8_obmc_sub_pixel_variance64x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x32 aom_highbd_8_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance64x64 aom_highbd_8_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x16 aom_highbd_8_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x32 aom_highbd_8_obmc_sub_pixel_variance8x32_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x4 aom_highbd_8_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_8_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_8_obmc_sub_pixel_variance8x8 aom_highbd_8_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_8_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_8_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x128 aom_highbd_8_sub_pixel_avg_variance128x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance128x64 aom_highbd_8_sub_pixel_avg_variance128x64_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x4 aom_highbd_8_sub_pixel_avg_variance16x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x64 aom_highbd_8_sub_pixel_avg_variance16x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x8 aom_highbd_8_sub_pixel_avg_variance32x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x128 aom_highbd_8_sub_pixel_avg_variance64x128_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x16 aom_highbd_8_sub_pixel_avg_variance64x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x32 aom_highbd_8_sub_pixel_avg_variance8x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x128 aom_highbd_8_sub_pixel_variance128x128_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance128x64 aom_highbd_8_sub_pixel_variance128x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x4 aom_highbd_8_sub_pixel_variance16x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x64 aom_highbd_8_sub_pixel_variance16x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x8 aom_highbd_8_sub_pixel_variance32x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x128 aom_highbd_8_sub_pixel_variance64x128_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x16 aom_highbd_8_sub_pixel_variance64x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x32 aom_highbd_8_sub_pixel_variance8x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_8_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x128 aom_highbd_8_variance128x128_sse2
+
+unsigned int aom_highbd_8_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance128x64 aom_highbd_8_variance128x64_sse2
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_sse2
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_sse2
+
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
+unsigned int aom_highbd_8_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x64 aom_highbd_8_variance16x64_sse2
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_sse2
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_sse2
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_sse2
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_sse2
+
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x8 aom_highbd_8_variance32x8_sse2
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x128 aom_highbd_8_variance64x128_sse2
+
+unsigned int aom_highbd_8_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x16 aom_highbd_8_variance64x16_sse2
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_sse2
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_sse2
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_sse2
+
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x32 aom_highbd_8_variance8x32_sse2
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+unsigned int aom_highbd_8_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_sse2
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *, int p);
+#define aom_highbd_avg_4x4 aom_highbd_avg_4x4_c
+
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *, int p);
+#define aom_highbd_avg_8x8 aom_highbd_avg_8x8_c
+
+void aom_highbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+void aom_highbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params, const int bd);
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+void aom_highbd_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h, int bd);
+
+void aom_highbd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+#define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
+
+void aom_highbd_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+RTCD_EXTERN void (*aom_highbd_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd);
+
+void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h);
+
+void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_sse2
+
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_sse2
+
+void aom_highbd_dc_128_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x4 aom_highbd_dc_128_predictor_16x4_c
+
+void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x64 aom_highbd_dc_128_predictor_16x64_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2
+
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2
+
+void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_sse2
+
+void aom_highbd_dc_128_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x64 aom_highbd_dc_128_predictor_32x64_c
+
+void aom_highbd_dc_128_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x8 aom_highbd_dc_128_predictor_32x8_c
+
+void aom_highbd_dc_128_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x16 aom_highbd_dc_128_predictor_4x16_c
+
+void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_sse2
+
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_sse2
+
+void aom_highbd_dc_128_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x16 aom_highbd_dc_128_predictor_64x16_c
+
+void aom_highbd_dc_128_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x32 aom_highbd_dc_128_predictor_64x32_c
+
+void aom_highbd_dc_128_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_64x64 aom_highbd_dc_128_predictor_64x64_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_sse2
+
+void aom_highbd_dc_128_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x32 aom_highbd_dc_128_predictor_8x32_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_sse2
+
+void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_sse2
+
+void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_sse2
+
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_sse2
+
+void aom_highbd_dc_left_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x4 aom_highbd_dc_left_predictor_16x4_c
+
+void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x64 aom_highbd_dc_left_predictor_16x64_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2
+
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2
+
+void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_sse2
+
+void aom_highbd_dc_left_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x64 aom_highbd_dc_left_predictor_32x64_c
+
+void aom_highbd_dc_left_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x8 aom_highbd_dc_left_predictor_32x8_c
+
+void aom_highbd_dc_left_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x16 aom_highbd_dc_left_predictor_4x16_c
+
+void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_sse2
+
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_sse2
+
+void aom_highbd_dc_left_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x16 aom_highbd_dc_left_predictor_64x16_c
+
+void aom_highbd_dc_left_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x32 aom_highbd_dc_left_predictor_64x32_c
+
+void aom_highbd_dc_left_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_64x64 aom_highbd_dc_left_predictor_64x64_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_sse2
+
+void aom_highbd_dc_left_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x32 aom_highbd_dc_left_predictor_8x32_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_sse2
+
+void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_sse2
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_sse2
+
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_sse2
+
+void aom_highbd_dc_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x4 aom_highbd_dc_predictor_16x4_c
+
+void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x64 aom_highbd_dc_predictor_16x64_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2
+
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2
+
+void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_sse2
+
+void aom_highbd_dc_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x64 aom_highbd_dc_predictor_32x64_c
+
+void aom_highbd_dc_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x8 aom_highbd_dc_predictor_32x8_c
+
+void aom_highbd_dc_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x16 aom_highbd_dc_predictor_4x16_c
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_sse2
+
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_sse2
+
+void aom_highbd_dc_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x16 aom_highbd_dc_predictor_64x16_c
+
+void aom_highbd_dc_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x32 aom_highbd_dc_predictor_64x32_c
+
+void aom_highbd_dc_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_64x64 aom_highbd_dc_predictor_64x64_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_sse2
+
+void aom_highbd_dc_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x32 aom_highbd_dc_predictor_8x32_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_sse2
+
+void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_sse2
+
+void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_sse2
+
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_sse2
+
+void aom_highbd_dc_top_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x4 aom_highbd_dc_top_predictor_16x4_c
+
+void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x64 aom_highbd_dc_top_predictor_16x64_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2
+
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2
+
+void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_sse2
+
+void aom_highbd_dc_top_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x64 aom_highbd_dc_top_predictor_32x64_c
+
+void aom_highbd_dc_top_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x8 aom_highbd_dc_top_predictor_32x8_c
+
+void aom_highbd_dc_top_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x16 aom_highbd_dc_top_predictor_4x16_c
+
+void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_sse2
+
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_sse2
+
+void aom_highbd_dc_top_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x16 aom_highbd_dc_top_predictor_64x16_c
+
+void aom_highbd_dc_top_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x32 aom_highbd_dc_top_predictor_64x32_c
+
+void aom_highbd_dc_top_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_64x64 aom_highbd_dc_top_predictor_64x64_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_sse2
+
+void aom_highbd_dc_top_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x32 aom_highbd_dc_top_predictor_8x32_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_sse2
+
+void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_sse2
+
+void aom_highbd_dist_wtd_comp_avg_pred_c(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+void aom_highbd_dist_wtd_comp_avg_pred_sse2(uint8_t *comp_pred8, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param);
+#define aom_highbd_dist_wtd_comp_avg_pred aom_highbd_dist_wtd_comp_avg_pred_sse2
+
+unsigned int aom_highbd_dist_wtd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x128_avg aom_highbd_dist_wtd_sad128x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad128x64_avg aom_highbd_dist_wtd_sad128x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x16_avg aom_highbd_dist_wtd_sad16x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x32_avg aom_highbd_dist_wtd_sad16x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x4_avg aom_highbd_dist_wtd_sad16x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x64_avg aom_highbd_dist_wtd_sad16x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad16x8_avg aom_highbd_dist_wtd_sad16x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x16_avg aom_highbd_dist_wtd_sad32x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x32_avg aom_highbd_dist_wtd_sad32x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x64_avg aom_highbd_dist_wtd_sad32x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad32x8_avg aom_highbd_dist_wtd_sad32x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x16_avg aom_highbd_dist_wtd_sad4x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x4_avg aom_highbd_dist_wtd_sad4x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad4x8_avg aom_highbd_dist_wtd_sad4x8_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x128_avg aom_highbd_dist_wtd_sad64x128_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x16_avg aom_highbd_dist_wtd_sad64x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x32_avg aom_highbd_dist_wtd_sad64x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad64x64_avg aom_highbd_dist_wtd_sad64x64_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x16_avg aom_highbd_dist_wtd_sad8x16_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x32_avg aom_highbd_dist_wtd_sad8x32_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x4_avg aom_highbd_dist_wtd_sad8x4_avg_c
+
+unsigned int aom_highbd_dist_wtd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS* jcp_param);
+#define aom_highbd_dist_wtd_sad8x8_avg aom_highbd_dist_wtd_sad8x8_avg_c
+
+void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_sse2
+
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_sse2
+
+void aom_highbd_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x4 aom_highbd_h_predictor_16x4_c
+
+void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x64 aom_highbd_h_predictor_16x64_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2
+
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2
+
+void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_sse2
+
+void aom_highbd_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x64 aom_highbd_h_predictor_32x64_c
+
+void aom_highbd_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x8 aom_highbd_h_predictor_32x8_c
+
+void aom_highbd_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x16 aom_highbd_h_predictor_4x16_c
+
+void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_sse2
+
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_sse2
+
+void aom_highbd_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x16 aom_highbd_h_predictor_64x16_c
+
+void aom_highbd_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x32 aom_highbd_h_predictor_64x32_c
+
+void aom_highbd_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_64x64 aom_highbd_h_predictor_64x64_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_sse2
+
+void aom_highbd_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x32 aom_highbd_h_predictor_8x32_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_sse2
+
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_sse2
+
+void aom_highbd_hadamard_16x16_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_16x16)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_32x32_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_32x32)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_hadamard_8x8_c(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+RTCD_EXTERN void (*aom_highbd_hadamard_8x8)(const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff);
+
+void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_14 aom_highbd_lpf_horizontal_14_sse2
+
+void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+void aom_highbd_lpf_horizontal_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd);
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_sse2
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_6 aom_highbd_lpf_horizontal_6_sse2
+
+void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_6_dual aom_highbd_lpf_horizontal_6_dual_sse2
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_8 aom_highbd_lpf_horizontal_8_sse2
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_14 aom_highbd_lpf_vertical_14_sse2
+
+void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_14_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_4 aom_highbd_lpf_vertical_4_sse2
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_6 aom_highbd_lpf_vertical_6_sse2
+
+void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_6_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_6_dual aom_highbd_lpf_vertical_6_dual_sse2
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_8 aom_highbd_lpf_vertical_8_sse2
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+
+unsigned int aom_highbd_masked_sad128x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad128x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad128x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad128x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x128_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x128_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x128)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x32_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_avx2(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
+
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c
+
+void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
+void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
+
+void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c
+
+void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c
+
+void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c
+
+void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
+
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c
+
+void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c
+
+void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
+void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_32x32_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_32x32_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_highbd_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_64x64_adaptive aom_highbd_quantize_b_64x64_adaptive_sse2
+
+void aom_highbd_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_highbd_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_highbd_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_highbd_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x4_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x4_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x16 aom_highbd_sad4x16_sse2
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x16_avg aom_highbd_sad4x16_avg_sse2
+
+void aom_highbd_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x3d aom_highbd_sad4x16x3d_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x16x4d aom_highbd_sad4x16x4d_sse2
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x4 aom_highbd_sad4x4_sse2
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x4_avg aom_highbd_sad4x4_avg_sse2
+
+void aom_highbd_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x3d aom_highbd_sad4x4x3d_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x4x4d aom_highbd_sad4x4x4d_sse2
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x8 aom_highbd_sad4x8_sse2
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x8_avg aom_highbd_sad4x8_avg_sse2
+
+void aom_highbd_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x3d aom_highbd_sad4x8x3d_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad4x8x4d aom_highbd_sad4x8x4d_sse2
+
+unsigned int aom_highbd_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x16 aom_highbd_sad8x16_sse2
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x16_avg aom_highbd_sad8x16_avg_sse2
+
+void aom_highbd_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x3d aom_highbd_sad8x16x3d_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x16x4d aom_highbd_sad8x16x4d_sse2
+
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x32 aom_highbd_sad8x32_sse2
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x32_avg aom_highbd_sad8x32_avg_sse2
+
+void aom_highbd_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x3d aom_highbd_sad8x32x3d_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x32x4d aom_highbd_sad8x32x4d_sse2
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x4 aom_highbd_sad8x4_sse2
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x4_avg aom_highbd_sad8x4_avg_sse2
+
+void aom_highbd_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x3d aom_highbd_sad8x4x3d_c
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x4x4d aom_highbd_sad8x4x4d_sse2
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x8 aom_highbd_sad8x8_sse2
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x8_avg aom_highbd_sad8x8_avg_sse2
+
+void aom_highbd_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x3d aom_highbd_sad8x8x3d_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad8x8x4d aom_highbd_sad8x8x4d_sse2
+
+unsigned int aom_highbd_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_16x4 aom_highbd_sad_skip_16x4_c
+
+void aom_highbd_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_16x4x4d aom_highbd_sad_skip_16x4x4d_c
+
+unsigned int aom_highbd_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_32x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_32x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x16 aom_highbd_sad_skip_4x16_sse2
+
+void aom_highbd_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x16x4d aom_highbd_sad_skip_4x16x4d_sse2
+
+unsigned int aom_highbd_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x4 aom_highbd_sad_skip_4x4_c
+
+void aom_highbd_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x4x4d aom_highbd_sad_skip_4x4x4d_c
+
+unsigned int aom_highbd_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_4x8 aom_highbd_sad_skip_4x8_sse2
+
+void aom_highbd_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_4x8x4d aom_highbd_sad_skip_4x8x4d_sse2
+
+unsigned int aom_highbd_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_highbd_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_highbd_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_highbd_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x16 aom_highbd_sad_skip_8x16_sse2
+
+void aom_highbd_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x16x4d aom_highbd_sad_skip_8x16x4d_sse2
+
+unsigned int aom_highbd_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x32 aom_highbd_sad_skip_8x32_sse2
+
+void aom_highbd_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x32x4d aom_highbd_sad_skip_8x32x4d_sse2
+
+unsigned int aom_highbd_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x4 aom_highbd_sad_skip_8x4_c
+
+void aom_highbd_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x4x4d aom_highbd_sad_skip_8x4x4d_c
+
+unsigned int aom_highbd_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad_skip_8x8 aom_highbd_sad_skip_8x8_sse2
+
+void aom_highbd_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_highbd_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_highbd_sad_skip_8x8x4d aom_highbd_sad_skip_8x8x4d_sse2
+
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c
+
+void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c
+
+void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c
+
+void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c
+
+void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c
+
+void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
+void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
+
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c
+
+void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
+void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
+
+void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c
+
+void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c
+
+void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c
+
+void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
+
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c
+
+void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c
+
+void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
+void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
+
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c
+
+void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c
+
+void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c
+
+void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c
+
+void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c
+
+void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
+int64_t aom_highbd_sse_c(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_highbd_sse)(const uint8_t *a8, int a_stride, const uint8_t *b8,int b_stride, int width, int height);
+
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_highbd_ssim_parms_8x8 aom_highbd_ssim_parms_8x8_c
+
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define aom_highbd_subtract_block aom_highbd_subtract_block_sse2
+
+void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_sse2
+
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_sse2
+
+void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c
+
+void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2
+
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2
+
+void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_sse2
+
+void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c
+
+void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c
+
+void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_sse2
+
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_sse2
+
+void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c
+
+void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c
+
+void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_sse2
+
+void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_sse2
+
+void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_sse2
+
+void aom_ifft16x16_float_c(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft16x16_float)(const float *input, float *temp, float *output);
+
+void aom_ifft2x2_float_c(const float *input, float *temp, float *output);
+#define aom_ifft2x2_float aom_ifft2x2_float_c
+
+void aom_ifft32x32_float_c(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft32x32_float)(const float *input, float *temp, float *output);
+
+void aom_ifft4x4_float_c(const float *input, float *temp, float *output);
+void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output);
+#define aom_ifft4x4_float aom_ifft4x4_float_sse2
+
+void aom_ifft8x8_float_c(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output);
+void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output);
+RTCD_EXTERN void (*aom_ifft8x8_float)(const float *input, float *temp, float *output);
+
+void aom_int_pro_col_c(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_col_avx2(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_col)(int16_t *vbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+void aom_int_pro_row_avx2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int width, const int height, int norm_factor);
+
+void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params);
+
+void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_14 aom_lpf_horizontal_14_sse2
+
+void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2
+
+void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2
+
+void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2
+
+void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2
+
+void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2
+
+void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2
+
+void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2
+
+void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2
+
+void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2
+
+void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2
+
+void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2
+
+void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2
+
+void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2
+
+void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0);
+#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2
+
+unsigned int aom_masked_sad128x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad128x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad128x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad128x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad128x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad128x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad128x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad16x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad16x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad16x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad32x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad32x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad32x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad4x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad4x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad4x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x128_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x128_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x128)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x128x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x128x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x128x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad64x64x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad64x64x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad64x64x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x16x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x16x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x16x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x32_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x32x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x32x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x32x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x4x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x4x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x4x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_avx2(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_masked_sad8x8x4d_c(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+void aom_masked_sad8x8x4d_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+RTCD_EXTERN void (*aom_masked_sad8x8x4d)(const uint8_t *src, int src_stride, const uint8_t *ref[4], int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned sads[4]);
+
+unsigned int aom_masked_sub_pixel_variance128x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance128x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance128x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance128x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x128_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x128_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x128)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_minmax_8x8 aom_minmax_8x8_sse2
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_sse2
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_sse2
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_sse2
+
+uint64_t aom_mse_16xh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+uint64_t aom_mse_16xh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_16xh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_c(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_sse2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_avx2(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit)(uint8_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+uint64_t aom_mse_wxh_16bit_highbd_c(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_sse2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+uint64_t aom_mse_wxh_16bit_highbd_avx2(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+RTCD_EXTERN uint64_t (*aom_mse_wxh_16bit_highbd)(uint16_t *dst, int dstride,uint16_t *src, int sstride, int w, int h);
+
+unsigned int aom_obmc_sad128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance128x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance128x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance128x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance16x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance32x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance4x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x128_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x128)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance64x64)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x16)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x32)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x4)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_sub_pixel_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_sub_pixel_variance8x8)(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance128x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance128x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance128x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x128_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x128_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x128)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x32_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_avx2(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32_adaptive aom_quantize_b_32x32_adaptive_sse2
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_64x64_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64_adaptive aom_quantize_b_64x64_adaptive_sse2
+
+void aom_quantize_b_adaptive_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_adaptive_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_adaptive)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+unsigned int aom_sad128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad128x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad128x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad128x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x16 aom_sad16x16_sse2
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_sse2
+
+void aom_sad16x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_sse2
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_sse2
+
+void aom_sad16x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x4 aom_sad16x4_sse2
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x4_avg aom_sad16x4_avg_sse2
+
+void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad16x4x3d aom_sad16x4x3d_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x64 aom_sad16x64_sse2
+
+unsigned int aom_sad16x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x64_avg aom_sad16x64_avg_sse2
+
+void aom_sad16x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x8 aom_sad16x8_sse2
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_sse2
+
+void aom_sad16x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x8 aom_sad32x8_sse2
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x8_avg aom_sad32x8_avg_sse2
+
+void aom_sad32x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x16 aom_sad4x16_sse2
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x16_avg aom_sad4x16_avg_sse2
+
+void aom_sad4x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x3d aom_sad4x16x3d_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x16x4d aom_sad4x16x4d_sse2
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x4 aom_sad4x4_sse2
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_sse2
+
+void aom_sad4x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x3d aom_sad4x4x3d_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x4x4d aom_sad4x4x4d_sse2
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_sse2
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_sse2
+
+void aom_sad4x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x3d aom_sad4x8x3d_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad4x8x4d aom_sad4x8x4d_sse2
+
+unsigned int aom_sad64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x128_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x128_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x128x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x16 aom_sad64x16_sse2
+
+unsigned int aom_sad64x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x16_avg aom_sad64x16_avg_sse2
+
+void aom_sad64x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x16 aom_sad8x16_sse2
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_sse2
+
+void aom_sad8x16x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x3d aom_sad8x16x3d_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x16x4d aom_sad8x16x4d_sse2
+
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x32 aom_sad8x32_sse2
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x32_avg aom_sad8x32_avg_sse2
+
+void aom_sad8x32x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x3d aom_sad8x32x3d_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x32x4d aom_sad8x32x4d_sse2
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_sse2
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_sse2
+
+void aom_sad8x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x3d aom_sad8x4x3d_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x4x4d aom_sad8x4x4d_sse2
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x8 aom_sad8x8_sse2
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_sse2
+
+void aom_sad8x8x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x3d aom_sad8x8x3d_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad8x8x4d aom_sad8x8x4d_sse2
+
+unsigned int aom_sad_skip_128x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_128x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_128x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_128x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_128x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_128x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x16 aom_sad_skip_16x16_sse2
+
+void aom_sad_skip_16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x32 aom_sad_skip_16x32_sse2
+
+void aom_sad_skip_16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x4 aom_sad_skip_16x4_c
+
+void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+
+unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x64 aom_sad_skip_16x64_sse2
+
+void aom_sad_skip_16x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_16x8 aom_sad_skip_16x8_sse2
+
+void aom_sad_skip_16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_32x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_32x8 aom_sad_skip_32x8_sse2
+
+void aom_sad_skip_32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_32x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_32x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x16 aom_sad_skip_4x16_sse2
+
+void aom_sad_skip_4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x16x4d aom_sad_skip_4x16x4d_sse2
+
+unsigned int aom_sad_skip_4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x4 aom_sad_skip_4x4_c
+
+void aom_sad_skip_4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x4x4d aom_sad_skip_4x4x4d_c
+
+unsigned int aom_sad_skip_4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_4x8 aom_sad_skip_4x8_sse2
+
+void aom_sad_skip_4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_4x8x4d aom_sad_skip_4x8x4d_sse2
+
+unsigned int aom_sad_skip_64x128_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x128)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x128x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x128x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x128x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_64x16 aom_sad_skip_64x16_sse2
+
+void aom_sad_skip_64x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad_skip_64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+void aom_sad_skip_64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+
+unsigned int aom_sad_skip_8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x16 aom_sad_skip_8x16_sse2
+
+void aom_sad_skip_8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x16x4d aom_sad_skip_8x16x4d_sse2
+
+unsigned int aom_sad_skip_8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x32 aom_sad_skip_8x32_sse2
+
+void aom_sad_skip_8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x32x4d aom_sad_skip_8x32x4d_sse2
+
+unsigned int aom_sad_skip_8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x4 aom_sad_skip_8x4_c
+
+void aom_sad_skip_8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x4x4d aom_sad_skip_8x4x4d_c
+
+unsigned int aom_sad_skip_8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad_skip_8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad_skip_8x8 aom_sad_skip_8x8_sse2
+
+void aom_sad_skip_8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+void aom_sad_skip_8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+#define aom_sad_skip_8x8x4d aom_sad_skip_8x8x4d_sse2
+
+int aom_satd_c(const tran_low_t *coeff, int length);
+int aom_satd_sse2(const tran_low_t *coeff, int length);
+int aom_satd_avx2(const tran_low_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const tran_low_t *coeff, int length);
+
+int aom_satd_lp_c(const int16_t *coeff, int length);
+int aom_satd_lp_sse2(const int16_t *coeff, int length);
+int aom_satd_lp_avx2(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd_lp)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
+
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_smooth_v_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+int64_t aom_sse_c(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+RTCD_EXTERN int64_t (*aom_sse)(const uint8_t *a, int a_stride, const uint8_t *b,int b_stride, int width, int height);
+
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+void aom_ssim_parms_8x8_sse2(const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr);
+#define aom_ssim_parms_8x8 aom_ssim_parms_8x8_sse2
+
+uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t N);
+#define aom_sum_squares_i16 aom_sum_squares_i16_sse2
+
+uint64_t aom_sum_sse_2d_i16_c(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width, int height, int *sum);
+uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width, int height, int *sum);
+RTCD_EXTERN uint64_t (*aom_sum_sse_2d_i16)(const int16_t *src, int src_stride, int width, int height, int *sum);
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x16 aom_v_predictor_16x16_sse2
+
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
+
+void aom_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x4 aom_v_predictor_16x4_sse2
+
+void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x64 aom_v_predictor_16x64_sse2
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
+
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x8 aom_v_predictor_32x8_sse2
+
+void aom_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x16 aom_v_predictor_4x16_sse2
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x4 aom_v_predictor_4x4_sse2
+
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
+
+void aom_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
+
+void aom_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x32 aom_v_predictor_8x32_sse2
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_sse2
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x8 aom_v_predictor_8x8_sse2
+
+uint64_t aom_var_2d_u16_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u16)(uint8_t *src, int src_stride, int width, int height);
+
+uint64_t aom_var_2d_u8_c(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width, int height);
+uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_var_2d_u8)(uint8_t *src, int src_stride, int width, int height);
+
+unsigned int aom_variance128x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance128x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance128x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x8_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x16 aom_variance4x16_sse2
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x4 aom_variance4x4_sse2
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x8 aom_variance4x8_sse2
+
+unsigned int aom_variance64x128_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x128)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x16 aom_variance8x16_sse2
+
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x32 aom_variance8x32_sse2
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x4 aom_variance8x4_sse2
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x8 aom_variance8x8_sse2
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
+
+void aom_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_avg_8x8_quad = aom_avg_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_avg_8x8_quad = aom_avg_8x8_quad_avx2;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
+    aom_comp_avg_pred = aom_comp_avg_pred_c;
+    if (flags & HAS_AVX2) aom_comp_avg_pred = aom_comp_avg_pred_avx2;
+    aom_comp_mask_pred = aom_comp_mask_pred_c;
+    if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
+    if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_flow_at_point = aom_compute_flow_at_point_c;
+    if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_horiz = aom_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_horiz = aom_convolve8_horiz_avx2;
+    aom_convolve8_vert = aom_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
+    aom_convolve_copy = aom_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_avx2;
+    aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x64 = aom_dc_128_predictor_32x64_avx2;
+    aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x16 = aom_dc_128_predictor_64x16_avx2;
+    aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x32 = aom_dc_128_predictor_64x32_avx2;
+    aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_64x64 = aom_dc_128_predictor_64x64_avx2;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_avx2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_avx2;
+    aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x64 = aom_dc_left_predictor_32x64_avx2;
+    aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x16 = aom_dc_left_predictor_64x16_avx2;
+    aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x32 = aom_dc_left_predictor_64x32_avx2;
+    aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_64x64 = aom_dc_left_predictor_64x64_avx2;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_avx2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_avx2;
+    aom_dc_predictor_32x64 = aom_dc_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x64 = aom_dc_predictor_32x64_avx2;
+    aom_dc_predictor_64x16 = aom_dc_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x16 = aom_dc_predictor_64x16_avx2;
+    aom_dc_predictor_64x32 = aom_dc_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x32 = aom_dc_predictor_64x32_avx2;
+    aom_dc_predictor_64x64 = aom_dc_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_64x64 = aom_dc_predictor_64x64_avx2;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_avx2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_avx2;
+    aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x64 = aom_dc_top_predictor_32x64_avx2;
+    aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x16 = aom_dc_top_predictor_64x16_avx2;
+    aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x32 = aom_dc_top_predictor_64x32_avx2;
+    aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_64x64 = aom_dc_top_predictor_64x64_avx2;
+    aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_pred = aom_dist_wtd_comp_avg_pred_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x128 = aom_dist_wtd_sub_pixel_avg_variance128x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance128x64 = aom_dist_wtd_sub_pixel_avg_variance128x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x16 = aom_dist_wtd_sub_pixel_avg_variance16x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x32 = aom_dist_wtd_sub_pixel_avg_variance16x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x4 = aom_dist_wtd_sub_pixel_avg_variance16x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x64 = aom_dist_wtd_sub_pixel_avg_variance16x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance16x8 = aom_dist_wtd_sub_pixel_avg_variance16x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x16 = aom_dist_wtd_sub_pixel_avg_variance32x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x32 = aom_dist_wtd_sub_pixel_avg_variance32x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x64 = aom_dist_wtd_sub_pixel_avg_variance32x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance32x8 = aom_dist_wtd_sub_pixel_avg_variance32x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x16 = aom_dist_wtd_sub_pixel_avg_variance4x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x4 = aom_dist_wtd_sub_pixel_avg_variance4x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance4x8 = aom_dist_wtd_sub_pixel_avg_variance4x8_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x128 = aom_dist_wtd_sub_pixel_avg_variance64x128_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x16 = aom_dist_wtd_sub_pixel_avg_variance64x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x32 = aom_dist_wtd_sub_pixel_avg_variance64x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance64x64 = aom_dist_wtd_sub_pixel_avg_variance64x64_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x16 = aom_dist_wtd_sub_pixel_avg_variance8x16_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x32 = aom_dist_wtd_sub_pixel_avg_variance8x32_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x4 = aom_dist_wtd_sub_pixel_avg_variance8x4_ssse3;
+    aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_sub_pixel_avg_variance8x8 = aom_dist_wtd_sub_pixel_avg_variance8x8_ssse3;
+    aom_fft16x16_float = aom_fft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_fft16x16_float = aom_fft16x16_float_avx2;
+    aom_fft32x32_float = aom_fft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_fft32x32_float = aom_fft32x32_float_avx2;
+    aom_fft8x8_float = aom_fft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_fft8x8_float = aom_fft8x8_float_avx2;
+    aom_get_blk_sse_sum = aom_get_blk_sse_sum_sse2;
+    if (flags & HAS_AVX2) aom_get_blk_sse_sum = aom_get_blk_sse_sum_avx2;
+    aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_16x16_dual = aom_get_var_sse_sum_16x16_dual_avx2;
+    aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_sse2;
+    if (flags & HAS_AVX2) aom_get_var_sse_sum_8x8_quad = aom_get_var_sse_sum_8x8_quad_avx2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_avx2;
+    aom_hadamard_16x16 = aom_hadamard_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_16x16 = aom_hadamard_16x16_avx2;
+    aom_hadamard_32x32 = aom_hadamard_32x32_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_32x32 = aom_hadamard_32x32_avx2;
+    aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_16x16 = aom_hadamard_lp_16x16_avx2;
+    aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_sse2;
+    if (flags & HAS_AVX2) aom_hadamard_lp_8x8_dual = aom_hadamard_lp_8x8_dual_avx2;
+    aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x128 = aom_highbd_10_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance128x64 = aom_highbd_10_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x4 = aom_highbd_10_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x64 = aom_highbd_10_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x8 = aom_highbd_10_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x16 = aom_highbd_10_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x128 = aom_highbd_10_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x16 = aom_highbd_10_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x32 = aom_highbd_10_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
+    aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_sse4_1;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_sse4_1;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_sse4_1;
+    aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x4 = aom_highbd_10_obmc_variance16x4_sse4_1;
+    aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x64 = aom_highbd_10_obmc_variance16x64_sse4_1;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_sse4_1;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_sse4_1;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_sse4_1;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_sse4_1;
+    aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x8 = aom_highbd_10_obmc_variance32x8_sse4_1;
+    aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x16 = aom_highbd_10_obmc_variance4x16_sse4_1;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_sse4_1;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_sse4_1;
+    aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x128 = aom_highbd_10_obmc_variance64x128_sse4_1;
+    aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x16 = aom_highbd_10_obmc_variance64x16_sse4_1;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_sse4_1;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_sse4_1;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_sse4_1;
+    aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x32 = aom_highbd_10_obmc_variance8x32_sse4_1;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_sse4_1;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x128 = aom_highbd_10_sub_pixel_variance128x128_avx2;
+    aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance128x64 = aom_highbd_10_sub_pixel_variance128x64_avx2;
+    aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x16 = aom_highbd_10_sub_pixel_variance16x16_avx2;
+    aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x32 = aom_highbd_10_sub_pixel_variance16x32_avx2;
+    aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance16x8 = aom_highbd_10_sub_pixel_variance16x8_avx2;
+    aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x16 = aom_highbd_10_sub_pixel_variance32x16_avx2;
+    aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x32 = aom_highbd_10_sub_pixel_variance32x32_avx2;
+    aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance32x64 = aom_highbd_10_sub_pixel_variance32x64_avx2;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x128 = aom_highbd_10_sub_pixel_variance64x128_avx2;
+    aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x32 = aom_highbd_10_sub_pixel_variance64x32_avx2;
+    aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance64x64 = aom_highbd_10_sub_pixel_variance64x64_avx2;
+    aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x16 = aom_highbd_10_sub_pixel_variance8x16_avx2;
+    aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_sub_pixel_variance8x8 = aom_highbd_10_sub_pixel_variance8x8_avx2;
+    aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x128 = aom_highbd_10_variance128x128_avx2;
+    aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance128x64 = aom_highbd_10_variance128x64_avx2;
+    aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x16 = aom_highbd_10_variance16x16_avx2;
+    aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x32 = aom_highbd_10_variance16x32_avx2;
+    aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x64 = aom_highbd_10_variance16x64_avx2;
+    aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance16x8 = aom_highbd_10_variance16x8_avx2;
+    aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x16 = aom_highbd_10_variance32x16_avx2;
+    aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x32 = aom_highbd_10_variance32x32_avx2;
+    aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x64 = aom_highbd_10_variance32x64_avx2;
+    aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance32x8 = aom_highbd_10_variance32x8_avx2;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_sse4_1;
+    aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x128 = aom_highbd_10_variance64x128_avx2;
+    aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x16 = aom_highbd_10_variance64x16_avx2;
+    aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x32 = aom_highbd_10_variance64x32_avx2;
+    aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance64x64 = aom_highbd_10_variance64x64_avx2;
+    aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x16 = aom_highbd_10_variance8x16_avx2;
+    aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x32 = aom_highbd_10_variance8x32_avx2;
+    aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_variance8x8 = aom_highbd_10_variance8x8_avx2;
+    aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x128 = aom_highbd_12_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance128x64 = aom_highbd_12_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x4 = aom_highbd_12_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x64 = aom_highbd_12_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x8 = aom_highbd_12_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x16 = aom_highbd_12_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x128 = aom_highbd_12_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x16 = aom_highbd_12_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x32 = aom_highbd_12_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x128 = aom_highbd_12_obmc_variance128x128_sse4_1;
+    aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance128x64 = aom_highbd_12_obmc_variance128x64_sse4_1;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_sse4_1;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_sse4_1;
+    aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x4 = aom_highbd_12_obmc_variance16x4_sse4_1;
+    aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x64 = aom_highbd_12_obmc_variance16x64_sse4_1;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_sse4_1;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_sse4_1;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_sse4_1;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_sse4_1;
+    aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x8 = aom_highbd_12_obmc_variance32x8_sse4_1;
+    aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x16 = aom_highbd_12_obmc_variance4x16_sse4_1;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_sse4_1;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_sse4_1;
+    aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x128 = aom_highbd_12_obmc_variance64x128_sse4_1;
+    aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x16 = aom_highbd_12_obmc_variance64x16_sse4_1;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_sse4_1;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_sse4_1;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_sse4_1;
+    aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x32 = aom_highbd_12_obmc_variance8x32_sse4_1;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_sse4_1;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_sse4_1;
+    aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x128 = aom_highbd_8_masked_sub_pixel_variance128x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance128x64 = aom_highbd_8_masked_sub_pixel_variance128x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x4 = aom_highbd_8_masked_sub_pixel_variance16x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x64 = aom_highbd_8_masked_sub_pixel_variance16x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x8 = aom_highbd_8_masked_sub_pixel_variance32x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x16 = aom_highbd_8_masked_sub_pixel_variance4x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x128 = aom_highbd_8_masked_sub_pixel_variance64x128_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x16 = aom_highbd_8_masked_sub_pixel_variance64x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x32 = aom_highbd_8_masked_sub_pixel_variance8x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x128 = aom_highbd_8_obmc_variance128x128_sse4_1;
+    aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance128x64 = aom_highbd_8_obmc_variance128x64_sse4_1;
+    aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x16 = aom_highbd_8_obmc_variance16x16_sse4_1;
+    aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x32 = aom_highbd_8_obmc_variance16x32_sse4_1;
+    aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x4 = aom_highbd_8_obmc_variance16x4_sse4_1;
+    aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x64 = aom_highbd_8_obmc_variance16x64_sse4_1;
+    aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance16x8 = aom_highbd_8_obmc_variance16x8_sse4_1;
+    aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x16 = aom_highbd_8_obmc_variance32x16_sse4_1;
+    aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x32 = aom_highbd_8_obmc_variance32x32_sse4_1;
+    aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x64 = aom_highbd_8_obmc_variance32x64_sse4_1;
+    aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance32x8 = aom_highbd_8_obmc_variance32x8_sse4_1;
+    aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x16 = aom_highbd_8_obmc_variance4x16_sse4_1;
+    aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x4 = aom_highbd_8_obmc_variance4x4_sse4_1;
+    aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance4x8 = aom_highbd_8_obmc_variance4x8_sse4_1;
+    aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x128 = aom_highbd_8_obmc_variance64x128_sse4_1;
+    aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x16 = aom_highbd_8_obmc_variance64x16_sse4_1;
+    aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x32 = aom_highbd_8_obmc_variance64x32_sse4_1;
+    aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance64x64 = aom_highbd_8_obmc_variance64x64_sse4_1;
+    aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x16 = aom_highbd_8_obmc_variance8x16_sse4_1;
+    aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x32 = aom_highbd_8_obmc_variance8x32_sse4_1;
+    aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x4 = aom_highbd_8_obmc_variance8x4_sse4_1;
+    aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_obmc_variance8x8 = aom_highbd_8_obmc_variance8x8_sse4_1;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_sse4_1;
+    aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_blend_a64_d16_mask = aom_highbd_blend_a64_d16_mask_avx2;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_sse4_1;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_sse4_1;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
+    aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_sse2;
+    if (flags & HAS_AVX2) aom_highbd_comp_mask_pred = aom_highbd_comp_mask_pred_avx2;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_avx2;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_avx2;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_16x16 = aom_highbd_hadamard_16x16_avx2;
+    aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_32x32 = aom_highbd_hadamard_32x32_avx2;
+    aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_c;
+    if (flags & HAS_AVX2) aom_highbd_hadamard_8x8 = aom_highbd_hadamard_8x8_avx2;
+    aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_avx2;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_avx2;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_avx2;
+    aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_avx2;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_avx2;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
+    aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x128 = aom_highbd_masked_sad128x128_avx2;
+    aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad128x64 = aom_highbd_masked_sad128x64_avx2;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_avx2;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_avx2;
+    aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x4 = aom_highbd_masked_sad16x4_avx2;
+    aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x64 = aom_highbd_masked_sad16x64_avx2;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_avx2;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_avx2;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_avx2;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_avx2;
+    aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad32x8 = aom_highbd_masked_sad32x8_avx2;
+    aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x16 = aom_highbd_masked_sad4x16_avx2;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_avx2;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_avx2;
+    aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x128 = aom_highbd_masked_sad64x128_avx2;
+    aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x16 = aom_highbd_masked_sad64x16_avx2;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_avx2;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_avx2;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_avx2;
+    aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x32 = aom_highbd_masked_sad8x32_avx2;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_avx2;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_avx2;
+    aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x128 = aom_highbd_obmc_sad128x128_avx2;
+    aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad128x64 = aom_highbd_obmc_sad128x64_avx2;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_avx2;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_avx2;
+    aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x4 = aom_highbd_obmc_sad16x4_avx2;
+    aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x64 = aom_highbd_obmc_sad16x64_avx2;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_avx2;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_avx2;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_avx2;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_avx2;
+    aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad32x8 = aom_highbd_obmc_sad32x8_avx2;
+    aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x16 = aom_highbd_obmc_sad4x16_avx2;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_avx2;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_avx2;
+    aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x128 = aom_highbd_obmc_sad64x128_avx2;
+    aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x16 = aom_highbd_obmc_sad64x16_avx2;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_avx2;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_avx2;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_avx2;
+    aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x32 = aom_highbd_obmc_sad8x32_avx2;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_avx2;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_avx2;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b = aom_highbd_quantize_b_avx2;
+    aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32 = aom_highbd_quantize_b_32x32_avx2;
+    aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_32x32_adaptive = aom_highbd_quantize_b_32x32_adaptive_avx2;
+    aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_64x64 = aom_highbd_quantize_b_64x64_avx2;
+    aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b_adaptive = aom_highbd_quantize_b_adaptive_avx2;
+    aom_highbd_sad128x128 = aom_highbd_sad128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128 = aom_highbd_sad128x128_avx2;
+    aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128_avg = aom_highbd_sad128x128_avg_avx2;
+    aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x3d = aom_highbd_sad128x128x3d_avx2;
+    aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x128x4d = aom_highbd_sad128x128x4d_avx2;
+    aom_highbd_sad128x64 = aom_highbd_sad128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64 = aom_highbd_sad128x64_avx2;
+    aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64_avg = aom_highbd_sad128x64_avg_avx2;
+    aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x3d = aom_highbd_sad128x64x3d_avx2;
+    aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad128x64x4d = aom_highbd_sad128x64x4d_avx2;
+    aom_highbd_sad16x16 = aom_highbd_sad16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16 = aom_highbd_sad16x16_avx2;
+    aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_avx2;
+    aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x3d = aom_highbd_sad16x16x3d_avx2;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_avx2;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32 = aom_highbd_sad16x32_avx2;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_avx2;
+    aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x3d = aom_highbd_sad16x32x3d_avx2;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_avx2;
+    aom_highbd_sad16x4 = aom_highbd_sad16x4_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4 = aom_highbd_sad16x4_avx2;
+    aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4_avg = aom_highbd_sad16x4_avg_avx2;
+    aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x3d = aom_highbd_sad16x4x3d_avx2;
+    aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x4x4d = aom_highbd_sad16x4x4d_avx2;
+    aom_highbd_sad16x64 = aom_highbd_sad16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64 = aom_highbd_sad16x64_avx2;
+    aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64_avg = aom_highbd_sad16x64_avg_avx2;
+    aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x3d = aom_highbd_sad16x64x3d_avx2;
+    aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x64x4d = aom_highbd_sad16x64x4d_avx2;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8 = aom_highbd_sad16x8_avx2;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_avx2;
+    aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x3d = aom_highbd_sad16x8x3d_avx2;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_avx2;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16 = aom_highbd_sad32x16_avx2;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_avx2;
+    aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x3d = aom_highbd_sad32x16x3d_avx2;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_avx2;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32 = aom_highbd_sad32x32_avx2;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_avx2;
+    aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x3d = aom_highbd_sad32x32x3d_avx2;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_avx2;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64 = aom_highbd_sad32x64_avx2;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_avx2;
+    aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x3d = aom_highbd_sad32x64x3d_avx2;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_avx2;
+    aom_highbd_sad32x8 = aom_highbd_sad32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8 = aom_highbd_sad32x8_avx2;
+    aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8_avg = aom_highbd_sad32x8_avg_avx2;
+    aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x3d = aom_highbd_sad32x8x3d_avx2;
+    aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x8x4d = aom_highbd_sad32x8x4d_avx2;
+    aom_highbd_sad64x128 = aom_highbd_sad64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128 = aom_highbd_sad64x128_avx2;
+    aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128_avg = aom_highbd_sad64x128_avg_avx2;
+    aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x3d = aom_highbd_sad64x128x3d_avx2;
+    aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x128x4d = aom_highbd_sad64x128x4d_avx2;
+    aom_highbd_sad64x16 = aom_highbd_sad64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16 = aom_highbd_sad64x16_avx2;
+    aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16_avg = aom_highbd_sad64x16_avg_avx2;
+    aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x3d = aom_highbd_sad64x16x3d_avx2;
+    aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x16x4d = aom_highbd_sad64x16x4d_avx2;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32 = aom_highbd_sad64x32_avx2;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_avx2;
+    aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x3d = aom_highbd_sad64x32x3d_avx2;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_avx2;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64 = aom_highbd_sad64x64_avx2;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_avx2;
+    aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x3d = aom_highbd_sad64x64x3d_avx2;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_avx2;
+    aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128 = aom_highbd_sad_skip_128x128_avx2;
+    aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x128x4d = aom_highbd_sad_skip_128x128x4d_avx2;
+    aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64 = aom_highbd_sad_skip_128x64_avx2;
+    aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_128x64x4d = aom_highbd_sad_skip_128x64x4d_avx2;
+    aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16 = aom_highbd_sad_skip_16x16_avx2;
+    aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x16x4d = aom_highbd_sad_skip_16x16x4d_avx2;
+    aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32 = aom_highbd_sad_skip_16x32_avx2;
+    aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x32x4d = aom_highbd_sad_skip_16x32x4d_avx2;
+    aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64 = aom_highbd_sad_skip_16x64_avx2;
+    aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x64x4d = aom_highbd_sad_skip_16x64x4d_avx2;
+    aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8 = aom_highbd_sad_skip_16x8_avx2;
+    aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_16x8x4d = aom_highbd_sad_skip_16x8x4d_avx2;
+    aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16 = aom_highbd_sad_skip_32x16_avx2;
+    aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x16x4d = aom_highbd_sad_skip_32x16x4d_avx2;
+    aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32 = aom_highbd_sad_skip_32x32_avx2;
+    aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x32x4d = aom_highbd_sad_skip_32x32x4d_avx2;
+    aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64 = aom_highbd_sad_skip_32x64_avx2;
+    aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x64x4d = aom_highbd_sad_skip_32x64x4d_avx2;
+    aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8 = aom_highbd_sad_skip_32x8_avx2;
+    aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_32x8x4d = aom_highbd_sad_skip_32x8x4d_avx2;
+    aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128 = aom_highbd_sad_skip_64x128_avx2;
+    aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_c;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x128x4d = aom_highbd_sad_skip_64x128x4d_avx2;
+    aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16 = aom_highbd_sad_skip_64x16_avx2;
+    aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x16x4d = aom_highbd_sad_skip_64x16x4d_avx2;
+    aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32 = aom_highbd_sad_skip_64x32_avx2;
+    aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x32x4d = aom_highbd_sad_skip_64x32x4d_avx2;
+    aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64 = aom_highbd_sad_skip_64x64_avx2;
+    aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad_skip_64x64x4d = aom_highbd_sad_skip_64x64x4d_avx2;
+    aom_highbd_sse = aom_highbd_sse_c;
+    if (flags & HAS_SSE4_1) aom_highbd_sse = aom_highbd_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_highbd_sse = aom_highbd_sse_avx2;
+    aom_ifft16x16_float = aom_ifft16x16_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft16x16_float = aom_ifft16x16_float_avx2;
+    aom_ifft32x32_float = aom_ifft32x32_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft32x32_float = aom_ifft32x32_float_avx2;
+    aom_ifft8x8_float = aom_ifft8x8_float_sse2;
+    if (flags & HAS_AVX2) aom_ifft8x8_float = aom_ifft8x8_float_avx2;
+    aom_int_pro_col = aom_int_pro_col_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_col = aom_int_pro_col_avx2;
+    aom_int_pro_row = aom_int_pro_row_sse2;
+    if (flags & HAS_AVX2) aom_int_pro_row = aom_int_pro_row_avx2;
+    aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
+    if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
+    aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2;
+    aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2;
+    aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2;
+    aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2;
+    if (flags & HAS_AVX2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_avx2;
+    aom_masked_sad128x128 = aom_masked_sad128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128 = aom_masked_sad128x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x128 = aom_masked_sad128x128_avx2;
+    aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x128x4d = aom_masked_sad128x128x4d_ssse3;
+    aom_masked_sad128x64 = aom_masked_sad128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64 = aom_masked_sad128x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad128x64 = aom_masked_sad128x64_avx2;
+    aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad128x64x4d = aom_masked_sad128x64x4d_ssse3;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x16 = aom_masked_sad16x16_avx2;
+    aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16x4d = aom_masked_sad16x16x4d_ssse3;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32 = aom_masked_sad16x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x32 = aom_masked_sad16x32_avx2;
+    aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32x4d = aom_masked_sad16x32x4d_ssse3;
+    aom_masked_sad16x4 = aom_masked_sad16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4 = aom_masked_sad16x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x4 = aom_masked_sad16x4_avx2;
+    aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x4x4d = aom_masked_sad16x4x4d_ssse3;
+    aom_masked_sad16x64 = aom_masked_sad16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64 = aom_masked_sad16x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x64 = aom_masked_sad16x64_avx2;
+    aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x64x4d = aom_masked_sad16x64x4d_ssse3;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8 = aom_masked_sad16x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad16x8 = aom_masked_sad16x8_avx2;
+    aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8x4d = aom_masked_sad16x8x4d_ssse3;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16 = aom_masked_sad32x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x16 = aom_masked_sad32x16_avx2;
+    aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16x4d = aom_masked_sad32x16x4d_ssse3;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32 = aom_masked_sad32x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x32 = aom_masked_sad32x32_avx2;
+    aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32x4d = aom_masked_sad32x32x4d_ssse3;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64 = aom_masked_sad32x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x64 = aom_masked_sad32x64_avx2;
+    aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64x4d = aom_masked_sad32x64x4d_ssse3;
+    aom_masked_sad32x8 = aom_masked_sad32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8 = aom_masked_sad32x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad32x8 = aom_masked_sad32x8_avx2;
+    aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x8x4d = aom_masked_sad32x8x4d_ssse3;
+    aom_masked_sad4x16 = aom_masked_sad4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16 = aom_masked_sad4x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x16 = aom_masked_sad4x16_avx2;
+    aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x16x4d = aom_masked_sad4x16x4d_ssse3;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4 = aom_masked_sad4x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x4 = aom_masked_sad4x4_avx2;
+    aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4x4d = aom_masked_sad4x4x4d_ssse3;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8 = aom_masked_sad4x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad4x8 = aom_masked_sad4x8_avx2;
+    aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8x4d = aom_masked_sad4x8x4d_ssse3;
+    aom_masked_sad64x128 = aom_masked_sad64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128 = aom_masked_sad64x128_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x128 = aom_masked_sad64x128_avx2;
+    aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x128x4d = aom_masked_sad64x128x4d_ssse3;
+    aom_masked_sad64x16 = aom_masked_sad64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16 = aom_masked_sad64x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x16 = aom_masked_sad64x16_avx2;
+    aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x16x4d = aom_masked_sad64x16x4d_ssse3;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32 = aom_masked_sad64x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x32 = aom_masked_sad64x32_avx2;
+    aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32x4d = aom_masked_sad64x32x4d_ssse3;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64 = aom_masked_sad64x64_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad64x64 = aom_masked_sad64x64_avx2;
+    aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64x4d = aom_masked_sad64x64x4d_ssse3;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16 = aom_masked_sad8x16_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x16 = aom_masked_sad8x16_avx2;
+    aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16x4d = aom_masked_sad8x16x4d_ssse3;
+    aom_masked_sad8x32 = aom_masked_sad8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32 = aom_masked_sad8x32_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x32 = aom_masked_sad8x32_avx2;
+    aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x32x4d = aom_masked_sad8x32x4d_ssse3;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4 = aom_masked_sad8x4_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x4 = aom_masked_sad8x4_avx2;
+    aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4x4d = aom_masked_sad8x4x4d_ssse3;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8 = aom_masked_sad8x8_ssse3;
+    if (flags & HAS_AVX2) aom_masked_sad8x8 = aom_masked_sad8x8_avx2;
+    aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8x4d = aom_masked_sad8x8x4d_ssse3;
+    aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x128 = aom_masked_sub_pixel_variance128x128_ssse3;
+    aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance128x64 = aom_masked_sub_pixel_variance128x64_ssse3;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_ssse3;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_ssse3;
+    aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x4 = aom_masked_sub_pixel_variance16x4_ssse3;
+    aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x64 = aom_masked_sub_pixel_variance16x64_ssse3;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_ssse3;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_ssse3;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_ssse3;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_ssse3;
+    aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x8 = aom_masked_sub_pixel_variance32x8_ssse3;
+    aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x16 = aom_masked_sub_pixel_variance4x16_ssse3;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_ssse3;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_ssse3;
+    aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x128 = aom_masked_sub_pixel_variance64x128_ssse3;
+    aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x16 = aom_masked_sub_pixel_variance64x16_ssse3;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_ssse3;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_ssse3;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_ssse3;
+    aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x32 = aom_masked_sub_pixel_variance8x32_ssse3;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_ssse3;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_ssse3;
+    aom_mse16x16 = aom_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
+    aom_mse_16xh_16bit = aom_mse_16xh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_16xh_16bit = aom_mse_16xh_16bit_avx2;
+    aom_mse_wxh_16bit = aom_mse_wxh_16bit_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit = aom_mse_wxh_16bit_avx2;
+    aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_sse2;
+    if (flags & HAS_AVX2) aom_mse_wxh_16bit_highbd = aom_mse_wxh_16bit_highbd_avx2;
+    aom_obmc_sad128x128 = aom_obmc_sad128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x128 = aom_obmc_sad128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x128 = aom_obmc_sad128x128_avx2;
+    aom_obmc_sad128x64 = aom_obmc_sad128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad128x64 = aom_obmc_sad128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad128x64 = aom_obmc_sad128x64_avx2;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x16 = aom_obmc_sad16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x16 = aom_obmc_sad16x16_avx2;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x32 = aom_obmc_sad16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x32 = aom_obmc_sad16x32_avx2;
+    aom_obmc_sad16x4 = aom_obmc_sad16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x4 = aom_obmc_sad16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x4 = aom_obmc_sad16x4_avx2;
+    aom_obmc_sad16x64 = aom_obmc_sad16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x64 = aom_obmc_sad16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x64 = aom_obmc_sad16x64_avx2;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x8 = aom_obmc_sad16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad16x8 = aom_obmc_sad16x8_avx2;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x16 = aom_obmc_sad32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x16 = aom_obmc_sad32x16_avx2;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x32 = aom_obmc_sad32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x32 = aom_obmc_sad32x32_avx2;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x64 = aom_obmc_sad32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x64 = aom_obmc_sad32x64_avx2;
+    aom_obmc_sad32x8 = aom_obmc_sad32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x8 = aom_obmc_sad32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad32x8 = aom_obmc_sad32x8_avx2;
+    aom_obmc_sad4x16 = aom_obmc_sad4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x16 = aom_obmc_sad4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x16 = aom_obmc_sad4x16_avx2;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x4 = aom_obmc_sad4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x4 = aom_obmc_sad4x4_avx2;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x8 = aom_obmc_sad4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad4x8 = aom_obmc_sad4x8_avx2;
+    aom_obmc_sad64x128 = aom_obmc_sad64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x128 = aom_obmc_sad64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x128 = aom_obmc_sad64x128_avx2;
+    aom_obmc_sad64x16 = aom_obmc_sad64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x16 = aom_obmc_sad64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x16 = aom_obmc_sad64x16_avx2;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x32 = aom_obmc_sad64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x32 = aom_obmc_sad64x32_avx2;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x64 = aom_obmc_sad64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad64x64 = aom_obmc_sad64x64_avx2;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x16 = aom_obmc_sad8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x16 = aom_obmc_sad8x16_avx2;
+    aom_obmc_sad8x32 = aom_obmc_sad8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x32 = aom_obmc_sad8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x32 = aom_obmc_sad8x32_avx2;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x4 = aom_obmc_sad8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x4 = aom_obmc_sad8x4_avx2;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x8 = aom_obmc_sad8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_sad8x8 = aom_obmc_sad8x8_avx2;
+    aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x128 = aom_obmc_sub_pixel_variance128x128_sse4_1;
+    aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance128x64 = aom_obmc_sub_pixel_variance128x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x16 = aom_obmc_sub_pixel_variance16x16_sse4_1;
+    aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x32 = aom_obmc_sub_pixel_variance16x32_sse4_1;
+    aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x4 = aom_obmc_sub_pixel_variance16x4_sse4_1;
+    aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x64 = aom_obmc_sub_pixel_variance16x64_sse4_1;
+    aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance16x8 = aom_obmc_sub_pixel_variance16x8_sse4_1;
+    aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x16 = aom_obmc_sub_pixel_variance32x16_sse4_1;
+    aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x32 = aom_obmc_sub_pixel_variance32x32_sse4_1;
+    aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x64 = aom_obmc_sub_pixel_variance32x64_sse4_1;
+    aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance32x8 = aom_obmc_sub_pixel_variance32x8_sse4_1;
+    aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x16 = aom_obmc_sub_pixel_variance4x16_sse4_1;
+    aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x4 = aom_obmc_sub_pixel_variance4x4_sse4_1;
+    aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance4x8 = aom_obmc_sub_pixel_variance4x8_sse4_1;
+    aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x128 = aom_obmc_sub_pixel_variance64x128_sse4_1;
+    aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x16 = aom_obmc_sub_pixel_variance64x16_sse4_1;
+    aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x32 = aom_obmc_sub_pixel_variance64x32_sse4_1;
+    aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance64x64 = aom_obmc_sub_pixel_variance64x64_sse4_1;
+    aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x16 = aom_obmc_sub_pixel_variance8x16_sse4_1;
+    aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x32 = aom_obmc_sub_pixel_variance8x32_sse4_1;
+    aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x4 = aom_obmc_sub_pixel_variance8x4_sse4_1;
+    aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sub_pixel_variance8x8 = aom_obmc_sub_pixel_variance8x8_sse4_1;
+    aom_obmc_variance128x128 = aom_obmc_variance128x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x128 = aom_obmc_variance128x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x128 = aom_obmc_variance128x128_avx2;
+    aom_obmc_variance128x64 = aom_obmc_variance128x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance128x64 = aom_obmc_variance128x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance128x64 = aom_obmc_variance128x64_avx2;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x16 = aom_obmc_variance16x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x16 = aom_obmc_variance16x16_avx2;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x32 = aom_obmc_variance16x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x32 = aom_obmc_variance16x32_avx2;
+    aom_obmc_variance16x4 = aom_obmc_variance16x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x4 = aom_obmc_variance16x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x4 = aom_obmc_variance16x4_avx2;
+    aom_obmc_variance16x64 = aom_obmc_variance16x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x64 = aom_obmc_variance16x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x64 = aom_obmc_variance16x64_avx2;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x8 = aom_obmc_variance16x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance16x8 = aom_obmc_variance16x8_avx2;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x16 = aom_obmc_variance32x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x16 = aom_obmc_variance32x16_avx2;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x32 = aom_obmc_variance32x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x32 = aom_obmc_variance32x32_avx2;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x64 = aom_obmc_variance32x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x64 = aom_obmc_variance32x64_avx2;
+    aom_obmc_variance32x8 = aom_obmc_variance32x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x8 = aom_obmc_variance32x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance32x8 = aom_obmc_variance32x8_avx2;
+    aom_obmc_variance4x16 = aom_obmc_variance4x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x16 = aom_obmc_variance4x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x16 = aom_obmc_variance4x16_avx2;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x4 = aom_obmc_variance4x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x4 = aom_obmc_variance4x4_avx2;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x8 = aom_obmc_variance4x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance4x8 = aom_obmc_variance4x8_avx2;
+    aom_obmc_variance64x128 = aom_obmc_variance64x128_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x128 = aom_obmc_variance64x128_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x128 = aom_obmc_variance64x128_avx2;
+    aom_obmc_variance64x16 = aom_obmc_variance64x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x16 = aom_obmc_variance64x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x16 = aom_obmc_variance64x16_avx2;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x32 = aom_obmc_variance64x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x32 = aom_obmc_variance64x32_avx2;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x64 = aom_obmc_variance64x64_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance64x64 = aom_obmc_variance64x64_avx2;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x16 = aom_obmc_variance8x16_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x16 = aom_obmc_variance8x16_avx2;
+    aom_obmc_variance8x32 = aom_obmc_variance8x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x32 = aom_obmc_variance8x32_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x32 = aom_obmc_variance8x32_avx2;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x4 = aom_obmc_variance8x4_avx2;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    if (flags & HAS_AVX2) aom_obmc_variance8x8 = aom_obmc_variance8x8_avx2;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+    aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_ssse3;
+    aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_avx2;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_avx2;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_avx2;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_avx2;
+    aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_avx2;
+    aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_ssse3;
+    aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_ssse3;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+    aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_avx2;
+    aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_avx2;
+    aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_avx2;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+    aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_ssse3;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3;
+    aom_quantize_b = aom_quantize_b_sse2;
+    if (flags & HAS_SSSE3) aom_quantize_b = aom_quantize_b_ssse3;
+    if (flags & HAS_AVX) aom_quantize_b = aom_quantize_b_avx;
+    if (flags & HAS_AVX2) aom_quantize_b = aom_quantize_b_avx2;
+    aom_quantize_b_32x32 = aom_quantize_b_32x32_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_32x32 = aom_quantize_b_32x32_ssse3;
+    if (flags & HAS_AVX) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx;
+    if (flags & HAS_AVX2) aom_quantize_b_32x32 = aom_quantize_b_32x32_avx2;
+    aom_quantize_b_64x64 = aom_quantize_b_64x64_c;
+    if (flags & HAS_SSSE3) aom_quantize_b_64x64 = aom_quantize_b_64x64_ssse3;
+    if (flags & HAS_AVX2) aom_quantize_b_64x64 = aom_quantize_b_64x64_avx2;
+    aom_quantize_b_adaptive = aom_quantize_b_adaptive_sse2;
+    if (flags & HAS_AVX2) aom_quantize_b_adaptive = aom_quantize_b_adaptive_avx2;
+    aom_sad128x128 = aom_sad128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128 = aom_sad128x128_avx2;
+    aom_sad128x128_avg = aom_sad128x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128_avg = aom_sad128x128_avg_avx2;
+    aom_sad128x128x3d = aom_sad128x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x128x3d = aom_sad128x128x3d_avx2;
+    aom_sad128x128x4d = aom_sad128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x128x4d = aom_sad128x128x4d_avx2;
+    aom_sad128x64 = aom_sad128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64 = aom_sad128x64_avx2;
+    aom_sad128x64_avg = aom_sad128x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64_avg = aom_sad128x64_avg_avx2;
+    aom_sad128x64x3d = aom_sad128x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad128x64x3d = aom_sad128x64x3d_avx2;
+    aom_sad128x64x4d = aom_sad128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad128x64x4d = aom_sad128x64x4d_avx2;
+    aom_sad16x16x3d = aom_sad16x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x16x3d = aom_sad16x16x3d_avx2;
+    aom_sad16x16x4d = aom_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x16x4d = aom_sad16x16x4d_avx2;
+    aom_sad16x32x3d = aom_sad16x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
+    aom_sad16x32x4d = aom_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4x4d = aom_sad16x4x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
+    aom_sad16x64x3d = aom_sad16x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x64x3d = aom_sad16x64x3d_avx2;
+    aom_sad16x64x4d = aom_sad16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x64x4d = aom_sad16x64x4d_avx2;
+    aom_sad16x8x3d = aom_sad16x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x8x3d = aom_sad16x8x3d_avx2;
+    aom_sad16x8x4d = aom_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad16x8x4d = aom_sad16x8x4d_avx2;
+    aom_sad32x16 = aom_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16 = aom_sad32x16_avx2;
+    aom_sad32x16_avg = aom_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16_avg = aom_sad32x16_avg_avx2;
+    aom_sad32x16x3d = aom_sad32x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x16x3d = aom_sad32x16x3d_avx2;
+    aom_sad32x16x4d = aom_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16x4d = aom_sad32x16x4d_avx2;
+    aom_sad32x32 = aom_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32 = aom_sad32x32_avx2;
+    aom_sad32x32_avg = aom_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32_avg = aom_sad32x32_avg_avx2;
+    aom_sad32x32x3d = aom_sad32x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x32x3d = aom_sad32x32x3d_avx2;
+    aom_sad32x32x4d = aom_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32x4d = aom_sad32x32x4d_avx2;
+    aom_sad32x64 = aom_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64 = aom_sad32x64_avx2;
+    aom_sad32x64_avg = aom_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64_avg = aom_sad32x64_avg_avx2;
+    aom_sad32x64x3d = aom_sad32x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x64x3d = aom_sad32x64x3d_avx2;
+    aom_sad32x64x4d = aom_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64x4d = aom_sad32x64x4d_avx2;
+    aom_sad32x8x3d = aom_sad32x8x3d_c;
+    if (flags & HAS_AVX2) aom_sad32x8x3d = aom_sad32x8x3d_avx2;
+    aom_sad32x8x4d = aom_sad32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x8x4d = aom_sad32x8x4d_avx2;
+    aom_sad64x128 = aom_sad64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128 = aom_sad64x128_avx2;
+    aom_sad64x128_avg = aom_sad64x128_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128_avg = aom_sad64x128_avg_avx2;
+    aom_sad64x128x3d = aom_sad64x128x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x128x3d = aom_sad64x128x3d_avx2;
+    aom_sad64x128x4d = aom_sad64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x128x4d = aom_sad64x128x4d_avx2;
+    aom_sad64x16x3d = aom_sad64x16x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x16x3d = aom_sad64x16x3d_avx2;
+    aom_sad64x16x4d = aom_sad64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x16x4d = aom_sad64x16x4d_avx2;
+    aom_sad64x32 = aom_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32 = aom_sad64x32_avx2;
+    aom_sad64x32_avg = aom_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32_avg = aom_sad64x32_avg_avx2;
+    aom_sad64x32x3d = aom_sad64x32x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x32x3d = aom_sad64x32x3d_avx2;
+    aom_sad64x32x4d = aom_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32x4d = aom_sad64x32x4d_avx2;
+    aom_sad64x64 = aom_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64 = aom_sad64x64_avx2;
+    aom_sad64x64_avg = aom_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64_avg = aom_sad64x64_avg_avx2;
+    aom_sad64x64x3d = aom_sad64x64x3d_c;
+    if (flags & HAS_AVX2) aom_sad64x64x3d = aom_sad64x64x3d_avx2;
+    aom_sad64x64x4d = aom_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64x4d = aom_sad64x64x4d_avx2;
+    aom_sad_skip_128x128 = aom_sad_skip_128x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128 = aom_sad_skip_128x128_avx2;
+    aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x128x4d = aom_sad_skip_128x128x4d_avx2;
+    aom_sad_skip_128x64 = aom_sad_skip_128x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64 = aom_sad_skip_128x64_avx2;
+    aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_128x64x4d = aom_sad_skip_128x64x4d_avx2;
+    aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
+    aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
+    aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_avx2;
+    aom_sad_skip_32x16 = aom_sad_skip_32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16 = aom_sad_skip_32x16_avx2;
+    aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x16x4d = aom_sad_skip_32x16x4d_avx2;
+    aom_sad_skip_32x32 = aom_sad_skip_32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32 = aom_sad_skip_32x32_avx2;
+    aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x32x4d = aom_sad_skip_32x32x4d_avx2;
+    aom_sad_skip_32x64 = aom_sad_skip_32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64 = aom_sad_skip_32x64_avx2;
+    aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x64x4d = aom_sad_skip_32x64x4d_avx2;
+    aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_32x8x4d = aom_sad_skip_32x8x4d_avx2;
+    aom_sad_skip_64x128 = aom_sad_skip_64x128_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128 = aom_sad_skip_64x128_avx2;
+    aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x128x4d = aom_sad_skip_64x128x4d_avx2;
+    aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x16x4d = aom_sad_skip_64x16x4d_avx2;
+    aom_sad_skip_64x32 = aom_sad_skip_64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32 = aom_sad_skip_64x32_avx2;
+    aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x32x4d = aom_sad_skip_64x32x4d_avx2;
+    aom_sad_skip_64x64 = aom_sad_skip_64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64 = aom_sad_skip_64x64_avx2;
+    aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad_skip_64x64x4d = aom_sad_skip_64x64x4d_avx2;
+    aom_satd = aom_satd_sse2;
+    if (flags & HAS_AVX2) aom_satd = aom_satd_avx2;
+    aom_satd_lp = aom_satd_lp_sse2;
+    if (flags & HAS_AVX2) aom_satd_lp = aom_satd_lp_avx2;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3;
+    aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_ssse3;
+    aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_ssse3;
+    aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_ssse3;
+    aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_ssse3;
+    aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_ssse3;
+    aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_ssse3;
+    aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_ssse3;
+    aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_ssse3;
+    aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_ssse3;
+    aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_ssse3;
+    aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_ssse3;
+    aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_ssse3;
+    aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_ssse3;
+    aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_ssse3;
+    aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_ssse3;
+    aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_ssse3;
+    aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_ssse3;
+    aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_ssse3;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_ssse3;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+    aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_ssse3;
+    aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_ssse3;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_ssse3;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_ssse3;
+    aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_ssse3;
+    aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_ssse3;
+    aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_ssse3;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+    aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_ssse3;
+    aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_ssse3;
+    aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_ssse3;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+    aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_ssse3;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_ssse3;
+    aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_ssse3;
+    aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_ssse3;
+    aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_ssse3;
+    aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_ssse3;
+    aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_ssse3;
+    aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_ssse3;
+    aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_ssse3;
+    aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_ssse3;
+    aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_ssse3;
+    aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_ssse3;
+    aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_ssse3;
+    aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_ssse3;
+    aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_ssse3;
+    aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_ssse3;
+    aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_ssse3;
+    aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_ssse3;
+    aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_ssse3;
+    aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_ssse3;
+    aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_ssse3;
+    aom_sse = aom_sse_c;
+    if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
+    if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_subtract_block = aom_subtract_block_sse2;
+    if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_avx2;
+    aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_sse2;
+    if (flags & HAS_AVX2) aom_sum_sse_2d_i16 = aom_sum_sse_2d_i16_avx2;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x16 = aom_v_predictor_32x16_avx2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x32 = aom_v_predictor_32x32_avx2;
+    aom_v_predictor_32x64 = aom_v_predictor_32x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x64 = aom_v_predictor_32x64_avx2;
+    aom_v_predictor_64x16 = aom_v_predictor_64x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x16 = aom_v_predictor_64x16_avx2;
+    aom_v_predictor_64x32 = aom_v_predictor_64x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x32 = aom_v_predictor_64x32_avx2;
+    aom_v_predictor_64x64 = aom_v_predictor_64x64_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_64x64 = aom_v_predictor_64x64_avx2;
+    aom_var_2d_u16 = aom_var_2d_u16_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u16 = aom_var_2d_u16_avx2;
+    aom_var_2d_u8 = aom_var_2d_u8_sse2;
+    if (flags & HAS_AVX2) aom_var_2d_u8 = aom_var_2d_u8_avx2;
+    aom_variance128x128 = aom_variance128x128_sse2;
+    if (flags & HAS_AVX2) aom_variance128x128 = aom_variance128x128_avx2;
+    aom_variance128x64 = aom_variance128x64_sse2;
+    if (flags & HAS_AVX2) aom_variance128x64 = aom_variance128x64_avx2;
+    aom_variance16x16 = aom_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
+    aom_variance16x32 = aom_variance16x32_sse2;
+    if (flags & HAS_AVX2) aom_variance16x32 = aom_variance16x32_avx2;
+    aom_variance16x4 = aom_variance16x4_sse2;
+    if (flags & HAS_AVX2) aom_variance16x4 = aom_variance16x4_avx2;
+    aom_variance16x64 = aom_variance16x64_sse2;
+    if (flags & HAS_AVX2) aom_variance16x64 = aom_variance16x64_avx2;
+    aom_variance16x8 = aom_variance16x8_sse2;
+    if (flags & HAS_AVX2) aom_variance16x8 = aom_variance16x8_avx2;
+    aom_variance32x16 = aom_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_variance32x16 = aom_variance32x16_avx2;
+    aom_variance32x32 = aom_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_variance32x32 = aom_variance32x32_avx2;
+    aom_variance32x64 = aom_variance32x64_sse2;
+    if (flags & HAS_AVX2) aom_variance32x64 = aom_variance32x64_avx2;
+    aom_variance32x8 = aom_variance32x8_sse2;
+    if (flags & HAS_AVX2) aom_variance32x8 = aom_variance32x8_avx2;
+    aom_variance64x128 = aom_variance64x128_sse2;
+    if (flags & HAS_AVX2) aom_variance64x128 = aom_variance64x128_avx2;
+    aom_variance64x16 = aom_variance64x16_sse2;
+    if (flags & HAS_AVX2) aom_variance64x16 = aom_variance64x16_avx2;
+    aom_variance64x32 = aom_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_variance64x32 = aom_variance64x32_avx2;
+    aom_variance64x64 = aom_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_variance64x64 = aom_variance64x64_avx2;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
+    if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
+    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
+    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/win/x64/config/aom_scale_rtcd.h b/media/libaom/config/win/x64/config/aom_scale_rtcd.h
new file mode 100644
index 0000000000..3b70fb47c3
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_scale_rtcd.h
@@ -0,0 +1,103 @@
+// This file is generated. Do not edit.
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
+#define aom_extend_frame_borders_plane_row aom_extend_frame_borders_plane_row_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, const int num_planes);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_yv12_partial_coloc_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_u aom_yv12_partial_coloc_copy_u_c
+
+void aom_yv12_partial_coloc_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_v aom_yv12_partial_coloc_copy_v_c
+
+void aom_yv12_partial_coloc_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int hstart, int hend, int vstart, int vend);
+#define aom_yv12_partial_coloc_copy_y aom_yv12_partial_coloc_copy_y_c
+
+void aom_yv12_partial_copy_u_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_bc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
+#define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
+
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+#define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
+
+void aom_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/media/libaom/config/win/x64/config/av1_rtcd.h b/media/libaom/config/win/x64/config/av1_rtcd.h
new file mode 100644
index 0000000000..b1cdc99700
--- /dev/null
+++ b/media/libaom/config/win/x64/config/av1_rtcd.h
@@ -0,0 +1,1217 @@
+// This file is generated. Do not edit.
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/odintrin.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/common.h"
+#include "av1/common/convolve.h"
+#include "av1/common/enums.h"
+#include "av1/common/filter.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct txfm_param;
+struct aom_variance_vtable;
+struct search_site_config;
+struct yv12_buffer_config;
+struct NN_CONFIG;
+typedef struct NN_CONFIG NN_CONFIG;
+
+enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);
+#if CONFIG_NN_V2
+enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);
+struct NN_CONFIG_V2;
+typedef struct NN_CONFIG_V2 NN_CONFIG_V2;
+struct FC_LAYER;
+typedef struct FC_LAYER FC_LAYER;
+#endif  // CONFIG_NN_V2
+
+struct CNN_CONFIG;
+typedef struct CNN_CONFIG CNN_CONFIG;
+struct CNN_LAYER_CONFIG;
+typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;
+struct CNN_THREAD_DATA;
+typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;
+struct CNN_BRANCH_CONFIG;
+typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;
+struct CNN_MULTI_OUT;
+typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+#endif
+
+typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
+
+typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
+                                   int dst_stride, int alpha_q3);
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+void aom_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                   int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                   int ref_stride, int subpel_search);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_sse2
+
+void aom_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_dist_wtd_comp_avg_upsampled_pred_ssse3(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+RTCD_EXTERN void (*aom_dist_wtd_comp_avg_upsampled_pred)(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                       const MV *const mv, uint8_t *comp_pred, const uint8_t *pred, int width,
+                                                       int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
+                                                       int ref_stride, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+
+void aom_highbd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                            const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                            int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_sse2
+
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+void aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                                const MV *const mv, uint8_t *comp_pred8, const uint8_t *pred8, int width,
+                                                                int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+                                                                int ref_stride, int bd, const DIST_WTD_COMP_PARAMS *jcp_param, int subpel_search);
+#define aom_highbd_dist_wtd_comp_avg_upsampled_pred aom_highbd_dist_wtd_comp_avg_upsampled_pred_sse2
+
+void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                                   const MV *const mv, uint8_t *comp_pred8, int width, int height, int subpel_x_q3,
+                                                   int subpel_y_q3, const uint8_t *ref8, int ref_stride, int bd, int subpel_search);
+#define aom_highbd_upsampled_pred aom_highbd_upsampled_pred_sse2
+
+void aom_quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr, const int log_scale);
+#define aom_quantize_b_helper aom_quantize_b_helper_c
+
+void aom_upsampled_pred_c(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+                                          const MV *const mv, uint8_t *comp_pred, int width, int height, int subpel_x_q3,
+                                          int subpel_y_q3, const uint8_t *ref, int ref_stride, int subpel_search);
+#define aom_upsampled_pred aom_upsampled_pred_sse2
+
+int av1_apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_sse4_1(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+int av1_apply_selfguided_restoration_avx2(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
+
+void av1_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_lp_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+int64_t av1_block_error_lp_avx2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+RTCD_EXTERN int64_t (*av1_block_error_lp)(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size);
+
+void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+
+void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+
+void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+void av1_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
+
+void av1_calc_indices_dim1_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim1_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim1)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_indices_dim2_c(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_sse2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+void av1_calc_indices_dim2_avx2(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+RTCD_EXTERN void (*av1_calc_indices_dim2)(const int16_t *data, const int16_t *centroids, uint8_t *indices, int64_t *total_dist, int n, int k);
+
+void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+void av1_calc_proj_params_high_bd_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+RTCD_EXTERN void (*av1_calc_proj_params_high_bd)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2], const sgr_params_type *params);
+
+void av1_cnn_activate_c(float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);
+#define av1_cnn_activate av1_cnn_activate_c
+
+void av1_cnn_add_c(float **input, int channels, int width, int height, int stride, const float **add);
+#define av1_cnn_add av1_cnn_add_c
+
+void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);
+#define av1_cnn_batchnorm av1_cnn_batchnorm_c
+
+void av1_cnn_convolve_no_maxpool_padding_valid_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+void av1_cnn_convolve_no_maxpool_padding_valid_avx2(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+RTCD_EXTERN void (*av1_cnn_convolve_no_maxpool_padding_valid)(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int cstep, int channel_step);
+
+void av1_cnn_deconvolve_c(const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);
+#define av1_cnn_deconvolve av1_cnn_deconvolve_c
+
+bool av1_cnn_predict_c(const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);
+#define av1_cnn_predict av1_cnn_predict_c
+
+void av1_compute_stats_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+RTCD_EXTERN void (*av1_compute_stats)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int16_t *dgd_avg, int16_t *src_avg, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, int use_downsampled_wiener_stats);
+
+void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_sse4_1(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+void av1_compute_stats_highbd_avx2(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+RTCD_EXTERN void (*av1_compute_stats_highbd)(int wiener_win, const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end, int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H, aom_bit_depth_t bit_depth);
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_convolve_2d_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+#define av1_convolve_2d_sr_intrabc av1_convolve_2d_sr_intrabc_c
+
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
+
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_convolve_x_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+#define av1_convolve_x_sr_intrabc av1_convolve_x_sr_intrabc_c
+
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+
+void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn);
+#define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
+
+void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params);
+
+void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params);
+
+void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
+
+void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
+
+void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
+
+double av1_estimate_noise_from_single_plane_c(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+double av1_estimate_noise_from_single_plane_avx2(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+RTCD_EXTERN double (*av1_estimate_noise_from_single_plane)(const uint8_t *src, int height, int width, int stride, int edge_thresh);
+
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_32x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_64x64_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x16_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_avx2(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void av1_fwht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*av1_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+uint32_t av1_get_crc32c_value_c(void *crc_calculator, uint8_t *p, size_t length);
+uint32_t av1_get_crc32c_value_sse4_2(void *crc_calculator, uint8_t *p, size_t length);
+RTCD_EXTERN uint32_t (*av1_get_crc32c_value)(void *crc_calculator, uint8_t *p, size_t length);
+
+void av1_get_horver_correlation_full_c(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_sse4_1(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+void av1_get_horver_correlation_full_avx2(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+RTCD_EXTERN void (*av1_get_horver_correlation_full)(const int16_t *diff, int stride, int w, int h, float *hcorr, float *vcorr);
+
+void av1_get_nz_map_contexts_c(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+void av1_get_nz_map_contexts_sse2(const uint8_t *const levels, const int16_t *const scan, const uint16_t eob, const TX_SIZE tx_size, const TX_CLASS tx_class, int8_t *const coeff_contexts);
+#define av1_get_nz_map_contexts av1_get_nz_map_contexts_sse2
+
+void av1_highbd_apply_temporal_filter_c(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_sse2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+void av1_highbd_apply_temporal_filter_avx2(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+RTCD_EXTERN void (*av1_highbd_apply_temporal_filter)(const struct yv12_buffer_config *frame_to_filter, const struct macroblockd *mbd, const BLOCK_SIZE block_size, const int mb_row, const int mb_col, const int num_planes, const double *noise_levels, const MV *subblock_mvs, const int *subblock_mses, const int q_factor, const int filter_strength, int tf_wgt_calc_lvl, const uint8_t *pred, uint32_t *accum, uint16_t *count);
+
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+RTCD_EXTERN int64_t (*av1_highbd_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_sse2
+
+void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_sse2
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_2d_sr_intrabc av1_highbd_convolve_2d_sr_intrabc_c
+
+void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
+
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_x_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+#define av1_highbd_convolve_x_sr_intrabc av1_highbd_convolve_x_sr_intrabc_c
+
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_intrabc_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd);
+#define av1_highbd_convolve_y_sr_intrabc av1_highbd_convolve_y_sr_intrabc_c
+
+void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
+
+void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+void av1_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+RTCD_EXTERN void (*av1_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
+
+double av1_highbd_estimate_noise_from_single_plane_c(const uint16_t *src, int height, int width, int stride, int bit_depth, int edge_thresh);
+#define av1_highbd_estimate_noise_from_single_plane av1_highbd_estimate_noise_from_single_plane_c
+
+void av1_highbd_filter_intra_edge_c(uint16_t *p, int sz, int strength);
+void av1_highbd_filter_intra_edge_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_highbd_filter_intra_edge)(uint16_t *p, int sz, int strength);
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c
+
+void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c
+
+void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c
+
+void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c
+
+void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c
+
+void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c
+
+void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+
+void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
+
+int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_highbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_highbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_sse4_1(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+void av1_highbd_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+RTCD_EXTERN void (*av1_highbd_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+
+void av1_highbd_upsample_intra_edge_c(uint16_t *p, int sz, int bd);
+void av1_highbd_upsample_intra_edge_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_highbd_upsample_intra_edge)(uint16_t *p, int sz, int bd);
+
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params, int bd);
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
+
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
+
+void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
+
+void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
+
+void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
+
+void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
+void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
+
+void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
+
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+
+void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
+
+int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_sse4_1(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+int64_t av1_lowbd_pixel_proj_error_avx2(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+RTCD_EXTERN int64_t (*av1_lowbd_pixel_proj_error)(const uint8_t *src8, int width, int height, int src_stride, const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params);
+
+void av1_nn_fast_softmax_16_c(const float *input_nodes, float *output);
+void av1_nn_fast_softmax_16_sse3(const float *input_nodes, float *output);
+RTCD_EXTERN void (*av1_nn_fast_softmax_16)(const float *input_nodes, float *output);
+
+void av1_nn_predict_c(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_sse3(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+void av1_nn_predict_avx2(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+RTCD_EXTERN void (*av1_nn_predict)(const float *input_nodes, const NN_CONFIG *const nn_config, int reduce_prec, float *const output);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp_64x64)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_lp_c(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_lp_avx2(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_lp)(const int16_t *coeff_ptr, intptr_t n_coeffs, const int16_t *round_ptr, const int16_t *quant_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);
+
+void av1_round_shift_array_c(int32_t *arr, int size, int bit);
+void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit);
+RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit);
+
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                int sgr_params_idx, int bit_depth, int highbd);
+
+void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+RTCD_EXTERN void (*av1_txb_init_levels)(const tran_low_t *const coeff, const int width, const int height, uint8_t *const levels);
+
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+void av1_wedge_compute_delta_squares_avx2(int16_t *d, const int16_t *a, const int16_t *b, int N);
+RTCD_EXTERN void (*av1_wedge_compute_delta_squares)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+
+int8_t av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int8_t av1_wedge_sign_from_residuals_avx2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+RTCD_EXTERN int8_t (*av1_wedge_sign_from_residuals)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+RTCD_EXTERN uint64_t (*av1_wedge_sse_from_residuals)(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+
+void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
+
+void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
+
+void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
+
+void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
+void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
+
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
+
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
+cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
+
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn cfl_get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*cfl_get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn cfl_get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*cfl_get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn cfl_get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn cfl_get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*cfl_get_subtract_average_fn)(TX_SIZE tx_size);
+
+void av1_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSSE3) aom_dist_wtd_comp_avg_upsampled_pred = aom_dist_wtd_comp_avg_upsampled_pred_ssse3;
+    av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_avx2;
+    av1_apply_temporal_filter = av1_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_apply_temporal_filter = av1_apply_temporal_filter_avx2;
+    av1_block_error = av1_block_error_sse2;
+    if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_block_error_lp = av1_block_error_lp_sse2;
+    if (flags & HAS_AVX2) av1_block_error_lp = av1_block_error_lp_avx2;
+    av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
+    av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
+    if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
+    av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
+    if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
+    av1_calc_indices_dim1 = av1_calc_indices_dim1_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim1 = av1_calc_indices_dim1_avx2;
+    av1_calc_indices_dim2 = av1_calc_indices_dim2_sse2;
+    if (flags & HAS_AVX2) av1_calc_indices_dim2 = av1_calc_indices_dim2_avx2;
+    av1_calc_proj_params = av1_calc_proj_params_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params = av1_calc_proj_params_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params = av1_calc_proj_params_avx2;
+    av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_c;
+    if (flags & HAS_SSE4_1) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_sse4_1;
+    if (flags & HAS_AVX2) av1_calc_proj_params_high_bd = av1_calc_proj_params_high_bd_avx2;
+    av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_c;
+    if (flags & HAS_AVX2) av1_cnn_convolve_no_maxpool_padding_valid = av1_cnn_convolve_no_maxpool_padding_valid_avx2;
+    av1_compute_stats = av1_compute_stats_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats = av1_compute_stats_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats = av1_compute_stats_avx2;
+    av1_compute_stats_highbd = av1_compute_stats_highbd_c;
+    if (flags & HAS_SSE4_1) av1_compute_stats_highbd = av1_compute_stats_highbd_sse4_1;
+    if (flags & HAS_AVX2) av1_compute_stats_highbd = av1_compute_stats_highbd_avx2;
+    av1_convolve_2d_scale = av1_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
+    av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
+    av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
+    av1_convolve_x_sr = av1_convolve_x_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
+    av1_convolve_y_sr = av1_convolve_y_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
+    av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_avx2;
+    av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_avx2;
+    av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2;
+    av1_dr_prediction_z1 = av1_dr_prediction_z1_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2;
+    av1_dr_prediction_z2 = av1_dr_prediction_z2_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2;
+    av1_dr_prediction_z3 = av1_dr_prediction_z3_c;
+    if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1;
+    if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2;
+    av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_c;
+    if (flags & HAS_AVX2) av1_estimate_noise_from_single_plane = av1_estimate_noise_from_single_plane_avx2;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
+    av1_filter_intra_predictor = av1_filter_intra_predictor_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_avx2;
+    av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x32 = av1_fwd_txfm2d_16x32_sse4_1;
+    av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x4 = av1_fwd_txfm2d_16x4_sse4_1;
+    av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x64 = av1_fwd_txfm2d_16x64_sse4_1;
+    av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_16x8 = av1_fwd_txfm2d_16x8_avx2;
+    av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x16 = av1_fwd_txfm2d_32x16_sse4_1;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_avx2;
+    av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x64 = av1_fwd_txfm2d_32x64_sse4_1;
+    av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x8 = av1_fwd_txfm2d_32x8_sse4_1;
+    av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x16 = av1_fwd_txfm2d_4x16_sse4_1;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
+    av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x8 = av1_fwd_txfm2d_4x8_sse4_1;
+    av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x16 = av1_fwd_txfm2d_64x16_sse4_1;
+    av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x32 = av1_fwd_txfm2d_64x32_sse4_1;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_avx2;
+    av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x16 = av1_fwd_txfm2d_8x16_avx2;
+    av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x32 = av1_fwd_txfm2d_8x32_sse4_1;
+    av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x4 = av1_fwd_txfm2d_8x4_sse4_1;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    if (flags & HAS_AVX2) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_avx2;
+    av1_fwht4x4 = av1_fwht4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwht4x4 = av1_fwht4x4_sse4_1;
+    av1_get_crc32c_value = av1_get_crc32c_value_c;
+    if (flags & HAS_SSE4_2) av1_get_crc32c_value = av1_get_crc32c_value_sse4_2;
+    av1_get_horver_correlation_full = av1_get_horver_correlation_full_c;
+    if (flags & HAS_SSE4_1) av1_get_horver_correlation_full = av1_get_horver_correlation_full_sse4_1;
+    if (flags & HAS_AVX2) av1_get_horver_correlation_full = av1_get_horver_correlation_full_avx2;
+    av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_sse2;
+    if (flags & HAS_AVX2) av1_highbd_apply_temporal_filter = av1_highbd_apply_temporal_filter_avx2;
+    av1_highbd_block_error = av1_highbd_block_error_sse2;
+    if (flags & HAS_AVX2) av1_highbd_block_error = av1_highbd_block_error_avx2;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
+    av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
+    av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
+    av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
+    av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d = av1_highbd_dist_wtd_convolve_2d_avx2;
+    av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_2d_copy = av1_highbd_dist_wtd_convolve_2d_copy_avx2;
+    av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_x = av1_highbd_dist_wtd_convolve_x_avx2;
+    av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_dist_wtd_convolve_y = av1_highbd_dist_wtd_convolve_y_avx2;
+    av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z1 = av1_highbd_dr_prediction_z1_avx2;
+    av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z2 = av1_highbd_dr_prediction_z2_avx2;
+    av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_c;
+    if (flags & HAS_AVX2) av1_highbd_dr_prediction_z3 = av1_highbd_dr_prediction_z3_avx2;
+    av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_filter_intra_edge = av1_highbd_filter_intra_edge_sse4_1;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_sse4_1;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_sse4_1;
+    av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
+    av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1;
+    av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_pixel_proj_error = av1_highbd_pixel_proj_error_avx2;
+    av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
+    if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
+    av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_highbd_upsample_intra_edge = av1_highbd_upsample_intra_edge_sse4_1;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2;
+    av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
+    if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
+    if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
+    av1_inv_txfm_add = av1_inv_txfm_add_c;
+    if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
+    if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
+    av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
+    if (flags & HAS_SSE4_1) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_sse4_1;
+    if (flags & HAS_AVX2) av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_avx2;
+    av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_c;
+    if (flags & HAS_SSE3) av1_nn_fast_softmax_16 = av1_nn_fast_softmax_16_sse3;
+    av1_nn_predict = av1_nn_predict_c;
+    if (flags & HAS_SSE3) av1_nn_predict = av1_nn_predict_sse3;
+    if (flags & HAS_AVX2) av1_nn_predict = av1_nn_predict_avx2;
+    av1_quantize_fp = av1_quantize_fp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
+    av1_quantize_fp_64x64 = av1_quantize_fp_64x64_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_64x64 = av1_quantize_fp_64x64_avx2;
+    av1_quantize_lp = av1_quantize_lp_sse2;
+    if (flags & HAS_AVX2) av1_quantize_lp = av1_quantize_lp_avx2;
+    av1_resize_and_extend_frame = av1_resize_and_extend_frame_c;
+    if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
+    av1_txb_init_levels = av1_txb_init_levels_c;
+    if (flags & HAS_SSE4_1) av1_txb_init_levels = av1_txb_init_levels_sse4_1;
+    if (flags & HAS_AVX2) av1_txb_init_levels = av1_txb_init_levels_avx2;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    if (flags & HAS_AVX2) av1_warp_affine = av1_warp_affine_avx2;
+    av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_sse2;
+    if (flags & HAS_AVX2) av1_wedge_compute_delta_squares = av1_wedge_compute_delta_squares_avx2;
+    av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sign_from_residuals = av1_wedge_sign_from_residuals_avx2;
+    av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_sse2;
+    if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
+    av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
+    if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
+    cdef_filter_16_0 = cdef_filter_16_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
+    cdef_filter_16_1 = cdef_filter_16_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
+    cdef_filter_16_2 = cdef_filter_16_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
+    cdef_filter_16_3 = cdef_filter_16_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
+    cdef_filter_8_0 = cdef_filter_8_0_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
+    cdef_filter_8_1 = cdef_filter_8_1_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
+    cdef_filter_8_2 = cdef_filter_8_2_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
+    cdef_filter_8_3 = cdef_filter_8_3_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
+    cdef_find_dir = cdef_find_dir_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
+    cdef_find_dir_dual = cdef_find_dir_dual_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
+    cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
+    cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
+    cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
+    cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
+    cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
+    cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
+    if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
+    if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
+    cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_hbd_fn = cfl_get_predict_hbd_fn_avx2;
+    cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) cfl_get_predict_lbd_fn = cfl_get_predict_lbd_fn_avx2;
+    cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) cfl_get_subtract_average_fn = cfl_get_subtract_average_fn_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif